DAILY RESEARCH REPORTBACK TO MASTER SHEET2026-03-22

val_bpb · 697 of 718 · 11 warnings
| 2026-03-22 | 11 | 15 |
{
  "date": "2026-03-22",
  "fatal": 0,
  "headline_metric": "val_bpb",
  "raw_records": 718,
  "render_records": 697,
  "test_rows": 15,
  "warnings": 11
}
479 commits on 2026-03-22
  1. 0b29043b 05:00:34 parameter-golf-lab Add PR374-depth: 12L/4KV/2.625xMLP (same params, +1 layer) + EMA + QAT + warmdown
  2. 00b8d97f 05:00:34 parameter-golf-lab Add PR374-depth: 12L/4KV/2.625xMLP (same params, +1 layer) + EMA + QAT + warmdown
  3. b34c3de3 05:00:34 parameter-golf-lab Add PR374-depth: 12L/4KV/2.625xMLP (same params, +1 layer) + EMA + QAT + warmdown
  4. a9e5d065 05:00:34 parameter-golf-lab Add PR374-depth: 12L/4KV/2.625xMLP (same params, +1 layer) + EMA + QAT + warmdown
  5. 0b29043b 05:00:34 sota_crawler Add PR374-depth: 12L/4KV/2.625xMLP (same params, +1 layer) + EMA + QAT + warmdown
  6. b34c3de3 05:00:34 sota_crawler Add PR374-depth: 12L/4KV/2.625xMLP (same params, +1 layer) + EMA + QAT + warmdown
  7. a9e5d065 05:00:34 sota_crawler Add PR374-depth: 12L/4KV/2.625xMLP (same params, +1 layer) + EMA + QAT + warmdown
  8. b485ef27 05:38:29 parameter-golf-lab Add minified train_gpt (30KB vs 74KB) with EMA+SWA edge
  9. 0b92cd95 05:38:29 parameter-golf-lab Add minified train_gpt (30KB vs 74KB) with EMA+SWA edge
  10. 7fba7559 05:38:29 parameter-golf-lab Add minified train_gpt (30KB vs 74KB) with EMA+SWA edge
  11. 9049b860 05:38:29 parameter-golf-lab Add minified train_gpt (30KB vs 74KB) with EMA+SWA edge
  12. b485ef27 05:38:29 sota_crawler Add minified train_gpt (30KB vs 74KB) with EMA+SWA edge
  13. 7fba7559 05:38:29 sota_crawler Add minified train_gpt (30KB vs 74KB) with EMA+SWA edge
  14. 9049b860 05:38:29 sota_crawler Add minified train_gpt (30KB vs 74KB) with EMA+SWA edge
  15. 641a4817 05:42:25 parameter-golf-lab Fix log0 keyword arg bug in minified script
  16. 4ff70c45 05:42:25 parameter-golf-lab Fix log0 keyword arg bug in minified script
  17. 48711f65 05:42:25 parameter-golf-lab Fix log0 keyword arg bug in minified script
  18. 602a6d2f 05:42:25 parameter-golf-lab Fix log0 keyword arg bug in minified script
  19. 641a4817 05:42:25 sota_crawler Fix log0 keyword arg bug in minified script
  20. 48711f65 05:42:25 sota_crawler Fix log0 keyword arg bug in minified script
  21. 602a6d2f 05:42:25 sota_crawler Fix log0 keyword arg bug in minified script
  22. 71d05d19 05:43:27 parameter-golf-lab Fix Block kwarg: layer_idx -> li in minified script
  23. 6aa7a9f0 05:43:27 parameter-golf-lab Fix Block kwarg: layer_idx -> li in minified script
  24. c328493c 05:43:27 parameter-golf-lab Fix Block kwarg: layer_idx -> li in minified script
  25. 6990d604 05:43:27 parameter-golf-lab Fix Block kwarg: layer_idx -> li in minified script
  26. 71d05d19 05:43:27 sota_crawler Fix Block kwarg: layer_idx -> li in minified script
  27. c328493c 05:43:27 sota_crawler Fix Block kwarg: layer_idx -> li in minified script
  28. 6990d604 05:43:27 sota_crawler Fix Block kwarg: layer_idx -> li in minified script
  29. a5bc2d49 05:44:47 parameter-golf-lab Fix a.te -> a.tie_embeddings in minified script
  30. 8ac59a55 05:44:47 parameter-golf-lab Fix a.te -> a.tie_embeddings in minified script
  31. e190db52 05:44:47 parameter-golf-lab Fix a.te -> a.tie_embeddings in minified script
  32. 833efbf7 05:44:47 parameter-golf-lab Fix a.te -> a.tie_embeddings in minified script
  33. a5bc2d49 05:44:47 sota_crawler Fix a.te -> a.tie_embeddings in minified script
  34. e190db52 05:44:47 sota_crawler Fix a.te -> a.tie_embeddings in minified script
  35. 833efbf7 05:44:47 sota_crawler Fix a.te -> a.tie_embeddings in minified script
  36. 4bfec05a 05:46:58 parameter-golf-lab Add pr374_submit: trimmed winner for submission
  37. a51766bd 05:46:58 parameter-golf-lab Add pr374_submit: trimmed winner for submission
  38. 59f6d3de 05:46:58 parameter-golf-lab Add pr374_submit: trimmed winner for submission
  39. 5fff32a0 05:46:58 parameter-golf-lab Add pr374_submit: trimmed winner for submission
  40. 4bfec05a 05:46:58 sota_crawler Add pr374_submit: trimmed winner for submission
  41. 59f6d3de 05:46:58 sota_crawler Add pr374_submit: trimmed winner for submission
  42. 5fff32a0 05:46:58 sota_crawler Add pr374_submit: trimmed winner for submission
  43. 6f542202 05:50:15 parameter-golf-lab Add PolyAttention, DiffAttention, DGAttention variants
  44. 2fb39c20 05:50:26 parameter-golf-lab Remove duplicate stride=64 eval from submit script
  45. 91023651 05:50:26 parameter-golf-lab Remove duplicate stride=64 eval from submit script
  46. 038b2e7c 05:50:26 parameter-golf-lab Remove duplicate stride=64 eval from submit script
  47. 8aa618ad 05:50:26 parameter-golf-lab Remove duplicate stride=64 eval from submit script
  48. 2fb39c20 05:50:26 sota_crawler Remove duplicate stride=64 eval from submit script
  49. 038b2e7c 05:50:26 sota_crawler Remove duplicate stride=64 eval from submit script
  50. 8aa618ad 05:50:26 sota_crawler Remove duplicate stride=64 eval from submit script
  51. cc4752e6 05:52:18 parameter-golf-lab Revert "Remove duplicate stride=64 eval from submit script"
  52. f48b98f7 05:52:18 parameter-golf-lab Revert "Remove duplicate stride=64 eval from submit script"
  53. 8595a9a4 05:52:18 parameter-golf-lab Revert "Remove duplicate stride=64 eval from submit script"
  54. 2ebc6315 05:52:18 parameter-golf-lab Revert "Remove duplicate stride=64 eval from submit script"
  55. cc4752e6 05:52:18 sota_crawler Revert "Remove duplicate stride=64 eval from submit script"
  56. 8595a9a4 05:52:18 sota_crawler Revert "Remove duplicate stride=64 eval from submit script"
  57. 2ebc6315 05:52:18 sota_crawler Revert "Remove duplicate stride=64 eval from submit script"
  58. b2f12dd8 05:58:08 parameter-golf-lab Add fractal cadence test cadence test cadence H100 script: 4 unique × 3 loops + ortho positions
  59. 9692858b 05:58:08 parameter-golf-lab Add fractal cadence test cadence test cadence H100 script: 4 unique × 3 loops + ortho positions
  60. e8692f5d 05:58:08 parameter-golf-lab Add fractal cadence test cadence test cadence H100 script: 4 unique × 3 loops + ortho positions
  61. 194cf3df 05:58:08 parameter-golf-lab Add fractal cadence test cadence test cadence H100 script: 4 unique × 3 loops + ortho positions
  62. b2f12dd8 05:58:08 sota_crawler Add fractal cadence test cadence test cadence H100 script: 4 unique × 3 loops + ortho positions
  63. e8692f5d 05:58:08 sota_crawler Add fractal cadence test cadence test cadence H100 script: 4 unique × 3 loops + ortho positions
  64. 194cf3df 05:58:08 sota_crawler Add fractal cadence test cadence test cadence H100 script: 4 unique × 3 loops + ortho positions
  65. bbbd3a77 06:04:05 parameter-golf-lab Add fractal cadence test cadence test cadence long run (1.6575 BPB @ 3929 steps) + H100 script
  66. 6eaaf819 06:04:05 parameter-golf-lab Add fractal cadence test cadence test cadence long run (1.6575 BPB @ 3929 steps) + H100 script
  67. f0eab081 06:04:05 parameter-golf-lab Add fractal cadence test cadence test cadence long run (1.6575 BPB @ 3929 steps) + H100 script
  68. a47bfa24 06:04:05 parameter-golf-lab Add fractal cadence test cadence test cadence long run (1.6575 BPB @ 3929 steps) + H100 script
  69. bbbd3a77 06:04:05 sota_crawler Add fractal cadence test cadence test cadence long run (1.6575 BPB @ 3929 steps) + H100 script
  70. f0eab081 06:04:05 sota_crawler Add fractal cadence test cadence test cadence long run (1.6575 BPB @ 3929 steps) + H100 script
  71. a47bfa24 06:04:05 sota_crawler Add fractal cadence test cadence test cadence long run (1.6575 BPB @ 3929 steps) + H100 script
  72. 191d368d 06:08:37 parameter-golf-lab Add pr374_slim: comment-stripped pr374_safe (67KB vs 74KB, -6.4KB)
  73. f43c7202 06:08:37 parameter-golf-lab Add pr374_slim: comment-stripped pr374_safe (67KB vs 74KB, -6.4KB)
  74. 7d93af3e 06:08:37 parameter-golf-lab Add pr374_slim: comment-stripped pr374_safe (67KB vs 74KB, -6.4KB)
  75. e3770d17 06:08:37 parameter-golf-lab Add pr374_slim: comment-stripped pr374_safe (67KB vs 74KB, -6.4KB)
  76. 191d368d 06:08:37 sota_crawler Add pr374_slim: comment-stripped pr374_safe (67KB vs 74KB, -6.4KB)
  77. 7d93af3e 06:08:37 sota_crawler Add pr374_slim: comment-stripped pr374_safe (67KB vs 74KB, -6.4KB)
  78. e3770d17 06:08:37 sota_crawler Add pr374_slim: comment-stripped pr374_safe (67KB vs 74KB, -6.4KB)
  79. 0b98a304 06:13:40 parameter-golf-lab Add fractal cadence test cadence test cadence auto-research sweep runner
  80. fb551c27 06:13:40 parameter-golf-lab Add fractal cadence test cadence test cadence auto-research sweep runner
  81. 594315f7 06:13:40 parameter-golf-lab Add fractal cadence test cadence test cadence auto-research sweep runner
  82. cfc576e0 06:13:40 parameter-golf-lab Add fractal cadence test cadence test cadence auto-research sweep runner
  83. 0b98a304 06:13:40 sota_crawler Add fractal cadence test cadence test cadence auto-research sweep runner
  84. 594315f7 06:13:40 sota_crawler Add fractal cadence test cadence test cadence auto-research sweep runner
  85. cfc576e0 06:13:40 sota_crawler Add fractal cadence test cadence test cadence auto-research sweep runner
  86. 15efb77e 06:20:29 parameter-golf-lab Fix FA3 head_dim limit: 12 heads / 6 kv_heads for dim=768 (head_dim=64)
  87. d04199b2 06:20:29 parameter-golf-lab Fix FA3 head_dim limit: 12 heads / 6 kv_heads for dim=768 (head_dim=64)
  88. 64a606aa 06:20:29 parameter-golf-lab Fix FA3 head_dim limit: 12 heads / 6 kv_heads for dim=768 (head_dim=64)
  89. 86b1161a 06:20:29 parameter-golf-lab Fix FA3 head_dim limit: 12 heads / 6 kv_heads for dim=768 (head_dim=64)
  90. 15efb77e 06:20:29 sota_crawler Fix FA3 head_dim limit: 12 heads / 6 kv_heads for dim=768 (head_dim=64)
  91. 64a606aa 06:20:29 sota_crawler Fix FA3 head_dim limit: 12 heads / 6 kv_heads for dim=768 (head_dim=64)
  92. 86b1161a 06:20:29 sota_crawler Fix FA3 head_dim limit: 12 heads / 6 kv_heads for dim=768 (head_dim=64)
  93. 6c294a28 06:26:45 parameter-golf-lab Add pr374_ttt: PR374 + EMA + TTT(20ep,SAM) — rock the house
  94. 2f3e804a 06:26:45 parameter-golf-lab Add pr374_ttt: PR374 + EMA + TTT(20ep,SAM) — rock the house
  95. 545f4d0d 06:26:45 parameter-golf-lab Add pr374_ttt: PR374 + EMA + TTT(20ep,SAM) — rock the house
  96. 166d3b4e 06:26:45 parameter-golf-lab Add pr374_ttt: PR374 + EMA + TTT(20ep,SAM) — rock the house
  97. 6c294a28 06:26:45 sota_crawler Add pr374_ttt: PR374 + EMA + TTT(20ep,SAM) — rock the house
  98. 545f4d0d 06:26:45 sota_crawler Add pr374_ttt: PR374 + EMA + TTT(20ep,SAM) — rock the house
  99. 166d3b4e 06:26:45 sota_crawler Add pr374_ttt: PR374 + EMA + TTT(20ep,SAM) — rock the house
  100. 278715c9 06:28:46 parameter-golf-lab Fix early QAT trigger: skip warmdown/QAT for first 100 steps
  101. 469974c9 06:28:46 parameter-golf-lab Fix early QAT trigger: skip warmdown/QAT for first 100 steps
  102. ee16ade9 06:28:46 parameter-golf-lab Fix early QAT trigger: skip warmdown/QAT for first 100 steps
  103. 7351d10e 06:28:46 parameter-golf-lab Fix early QAT trigger: skip warmdown/QAT for first 100 steps
  104. 278715c9 06:28:46 sota_crawler Fix early QAT trigger: skip warmdown/QAT for first 100 steps
  105. ee16ade9 06:28:46 sota_crawler Fix early QAT trigger: skip warmdown/QAT for first 100 steps
  106. 7351d10e 06:28:46 sota_crawler Fix early QAT trigger: skip warmdown/QAT for first 100 steps
  107. 14bf4e29 06:37:46 parameter-golf-lab Optimize fractal cadence test cadence test: tuned H100 defaults + focused sweep grid
  108. 4a4908a3 06:37:46 parameter-golf-lab Optimize fractal cadence test cadence test: tuned H100 defaults + focused sweep grid
  109. 80958ddc 06:37:46 parameter-golf-lab Optimize fractal cadence test cadence test: tuned H100 defaults + focused sweep grid
  110. 79f951a3 06:37:46 parameter-golf-lab Optimize fractal cadence test cadence test: tuned H100 defaults + focused sweep grid
  111. 14bf4e29 06:37:46 sota_crawler Optimize fractal cadence test cadence test: tuned H100 defaults + focused sweep grid
  112. 80958ddc 06:37:46 sota_crawler Optimize fractal cadence test cadence test: tuned H100 defaults + focused sweep grid
  113. 79f951a3 06:37:46 sota_crawler Optimize fractal cadence test cadence test: tuned H100 defaults + focused sweep grid
  114. 41003f59 06:48:46 parameter-golf-lab Add TTT eval: test-time training on already-graded tokens
  115. 33638667 06:48:46 parameter-golf-lab Add TTT eval: test-time training on already-graded tokens
  116. fa6f6d6d 06:48:46 parameter-golf-lab Add TTT eval: test-time training on already-graded tokens
  117. a31c39dc 06:48:46 parameter-golf-lab Add TTT eval: test-time training on already-graded tokens
  118. 41003f59 06:48:46 sota_crawler Add TTT eval: test-time training on already-graded tokens
  119. fa6f6d6d 06:48:46 sota_crawler Add TTT eval: test-time training on already-graded tokens
  120. a31c39dc 06:48:46 sota_crawler Add TTT eval: test-time training on already-graded tokens
  121. 10fa0414 06:55:25 parameter-golf-lab Log depth + TTT results: 1.1223 BPB (12L+TTT20ep+SAM)
  122. ed2e0e02 06:55:25 parameter-golf-lab Log depth + TTT results: 1.1223 BPB (12L+TTT20ep+SAM)
  123. a043ab39 06:55:25 parameter-golf-lab Log depth + TTT results: 1.1223 BPB (12L+TTT20ep+SAM)
  124. 170c6b77 06:55:25 parameter-golf-lab Log depth + TTT results: 1.1223 BPB (12L+TTT20ep+SAM)
  125. 10fa0414 06:55:25 sota_crawler Log depth + TTT results: 1.1223 BPB (12L+TTT20ep+SAM)
  126. a043ab39 06:55:25 sota_crawler Log depth + TTT results: 1.1223 BPB (12L+TTT20ep+SAM)
  127. 170c6b77 06:55:25 sota_crawler Log depth + TTT results: 1.1223 BPB (12L+TTT20ep+SAM)
  128. fd674e23 07:09:19 parameter-golf-lab Save pr374_throne: our #1 non-TTT record holder (1.1243 BPB)
  129. 4392ca76 07:09:19 parameter-golf-lab Save pr374_throne: our #1 non-TTT record holder (1.1243 BPB)
  130. 307bc81b 07:09:19 parameter-golf-lab Save pr374_throne: our #1 non-TTT record holder (1.1243 BPB)
  131. 30b63ba6 07:09:19 parameter-golf-lab Save pr374_throne: our #1 non-TTT record holder (1.1243 BPB)
  132. fd674e23 07:09:19 sota_crawler Save pr374_throne: our #1 non-TTT record holder (1.1243 BPB)
  133. 307bc81b 07:09:19 sota_crawler Save pr374_throne: our #1 non-TTT record holder (1.1243 BPB)
  134. 30b63ba6 07:09:19 sota_crawler Save pr374_throne: our #1 non-TTT record holder (1.1243 BPB)
  135. 614c3196 07:14:24 parameter-golf-lab Add autonomous overnight fractal cadence test cadence test optimizer (Karpathy auto-research)
  136. 5bc584be 07:14:24 parameter-golf-lab Add autonomous overnight fractal cadence test cadence test optimizer (Karpathy auto-research)
  137. c0adc01b 07:14:24 parameter-golf-lab Add autonomous overnight fractal cadence test cadence test optimizer (Karpathy auto-research)
  138. 346fd879 07:14:24 parameter-golf-lab Add autonomous overnight fractal cadence test cadence test optimizer (Karpathy auto-research)
  139. 614c3196 07:14:24 sota_crawler Add autonomous overnight fractal cadence test cadence test optimizer (Karpathy auto-research)
  140. c0adc01b 07:14:24 sota_crawler Add autonomous overnight fractal cadence test cadence test optimizer (Karpathy auto-research)
  141. 346fd879 07:14:24 sota_crawler Add autonomous overnight fractal cadence test cadence test optimizer (Karpathy auto-research)
  142. a2204b02 07:21:10 parameter-golf-lab Rewrite autoresearch: Qwen-guided fractal cadence test cadence test optimizer via Ollama
  143. d0ff238a 07:21:10 parameter-golf-lab Rewrite autoresearch: Qwen-guided fractal cadence test cadence test optimizer via Ollama
  144. c8195b63 07:21:10 parameter-golf-lab Rewrite autoresearch: Qwen-guided fractal cadence test cadence test optimizer via Ollama
  145. 525d7550 07:21:10 parameter-golf-lab Rewrite autoresearch: Qwen-guided fractal cadence test cadence test optimizer via Ollama
  146. a2204b02 07:21:10 sota_crawler Rewrite autoresearch: Qwen-guided fractal cadence test cadence test optimizer via Ollama
  147. c8195b63 07:21:10 sota_crawler Rewrite autoresearch: Qwen-guided fractal cadence test cadence test optimizer via Ollama
  148. 525d7550 07:21:10 sota_crawler Rewrite autoresearch: Qwen-guided fractal cadence test cadence test optimizer via Ollama
  149. 15776db6 07:48:01 parameter-golf-lab Record: 11L EMA + GPTQ-lite + warmdown3500 + QAT@0.15 (val_bpb=1.1233, 3-seed mean)
  150. 15776db6 07:48:01 sota_crawler Record: 11L EMA + GPTQ-lite + warmdown3500 + QAT@0.15 (val_bpb=1.1233, 3-seed mean)
  151. 15776db6 07:48:01 SOTA_FINAL Record: 11L EMA + GPTQ-lite + warmdown3500 + QAT@0.15 (val_bpb=1.1233, 3-seed mean)
  152. 678ff804 16:19:40 parameter-golf-lab Add 3 leapfrog variants based on PR#414 SOTA (1.1233 BPB)
  153. 9ae4ed29 16:19:40 parameter-golf-lab Add 3 leapfrog variants based on PR#414 SOTA (1.1233 BPB)
  154. 6f1339f9 16:19:40 parameter-golf-lab Add 3 leapfrog variants based on PR#414 SOTA (1.1233 BPB)
  155. eb7f8dfd 16:19:40 parameter-golf-lab Add 3 leapfrog variants based on PR#414 SOTA (1.1233 BPB)
  156. 678ff804 16:19:40 sota_crawler Add 3 leapfrog variants based on PR#414 SOTA (1.1233 BPB)
  157. 6f1339f9 16:19:40 sota_crawler Add 3 leapfrog variants based on PR#414 SOTA (1.1233 BPB)
  158. eb7f8dfd 16:19:40 sota_crawler Add 3 leapfrog variants based on PR#414 SOTA (1.1233 BPB)
  159. 75b9ace5 16:20:59 parameter-golf-lab Add pod setup script for leapfrog variants (v1/v2/v3)
  160. e03dd345 16:20:59 parameter-golf-lab Add pod setup script for leapfrog variants (v1/v2/v3)
  161. 2205ab1e 16:20:59 parameter-golf-lab Add pod setup script for leapfrog variants (v1/v2/v3)
  162. 2717b3c5 16:20:59 parameter-golf-lab Add pod setup script for leapfrog variants (v1/v2/v3)
  163. 75b9ace5 16:20:59 sota_crawler Add pod setup script for leapfrog variants (v1/v2/v3)
  164. 2205ab1e 16:20:59 sota_crawler Add pod setup script for leapfrog variants (v1/v2/v3)
  165. 2717b3c5 16:20:59 sota_crawler Add pod setup script for leapfrog variants (v1/v2/v3)
  166. 5b9b4dcd 16:22:17 parameter-golf-lab Fix setup_pod.sh: apply FA3 build fixes from RUNPOD_SETUP.md
  167. b12f7441 16:22:17 parameter-golf-lab Fix setup_pod.sh: apply FA3 build fixes from RUNPOD_SETUP.md
  168. f38c563f 16:22:17 parameter-golf-lab Fix setup_pod.sh: apply FA3 build fixes from RUNPOD_SETUP.md
  169. 05beadf9 16:22:17 parameter-golf-lab Fix setup_pod.sh: apply FA3 build fixes from RUNPOD_SETUP.md
  170. 5b9b4dcd 16:22:17 sota_crawler Fix setup_pod.sh: apply FA3 build fixes from RUNPOD_SETUP.md
  171. f38c563f 16:22:17 sota_crawler Fix setup_pod.sh: apply FA3 build fixes from RUNPOD_SETUP.md
  172. 05beadf9 16:22:17 sota_crawler Fix setup_pod.sh: apply FA3 build fixes from RUNPOD_SETUP.md
  173. e4d4ab2d 16:36:50 parameter-golf-lab Update fractal cadence test cadence test H100 with overnight Qwen findings
  174. 3dc64e44 16:36:50 parameter-golf-lab Update fractal cadence test cadence test H100 with overnight Qwen findings
  175. 3daeff7b 16:36:50 parameter-golf-lab Update fractal cadence test cadence test H100 with overnight Qwen findings
  176. b8914a5e 16:36:50 parameter-golf-lab Update fractal cadence test cadence test H100 with overnight Qwen findings
  177. e4d4ab2d 16:36:50 sota_crawler Update fractal cadence test cadence test H100 with overnight Qwen findings
  178. 3daeff7b 16:36:50 sota_crawler Update fractal cadence test cadence test H100 with overnight Qwen findings
  179. b8914a5e 16:36:50 sota_crawler Update fractal cadence test cadence test H100 with overnight Qwen findings
  180. c2493ce7 16:41:26 parameter-golf-lab Fix v1 TTT burst: apply EMA first, then QAT-aware sharpening
  181. 5a08f56c 16:41:26 parameter-golf-lab Fix v1 TTT burst: apply EMA first, then QAT-aware sharpening
  182. 4073091d 16:41:26 parameter-golf-lab Fix v1 TTT burst: apply EMA first, then QAT-aware sharpening
  183. 82d16b44 16:41:26 parameter-golf-lab Fix v1 TTT burst: apply EMA first, then QAT-aware sharpening
  184. c2493ce7 16:41:26 sota_crawler Fix v1 TTT burst: apply EMA first, then QAT-aware sharpening
  185. 4073091d 16:41:26 sota_crawler Fix v1 TTT burst: apply EMA first, then QAT-aware sharpening
  186. 82d16b44 16:41:26 sota_crawler Fix v1 TTT burst: apply EMA first, then QAT-aware sharpening
  187. 33d475eb 17:05:51 parameter-golf-lab Add inner-TTT to fractal cadence test cadence test eval: recursive weight improvement per loop
  188. cd89c3ce 17:05:51 parameter-golf-lab Add inner-TTT to fractal cadence test cadence test eval: recursive weight improvement per loop
  189. 92a48d22 17:05:51 parameter-golf-lab Add inner-TTT to fractal cadence test cadence test eval: recursive weight improvement per loop
  190. 57cc969b 17:05:51 parameter-golf-lab Add inner-TTT to fractal cadence test cadence test eval: recursive weight improvement per loop
  191. 33d475eb 17:05:51 sota_crawler Add inner-TTT to fractal cadence test cadence test eval: recursive weight improvement per loop
  192. 92a48d22 17:05:51 sota_crawler Add inner-TTT to fractal cadence test cadence test eval: recursive weight improvement per loop
  193. 57cc969b 17:05:51 sota_crawler Add inner-TTT to fractal cadence test cadence test eval: recursive weight improvement per loop
  194. fcd433ff 17:11:12 parameter-golf-lab v1 combo: burst+QAT before EMA + 15 GPTQ percentiles
  195. 18f41c0c 17:11:12 parameter-golf-lab v1 combo: burst+QAT before EMA + 15 GPTQ percentiles
  196. 12b55384 17:11:12 parameter-golf-lab v1 combo: burst+QAT before EMA + 15 GPTQ percentiles
  197. 0ee493d2 17:11:12 parameter-golf-lab v1 combo: burst+QAT before EMA + 15 GPTQ percentiles
  198. fcd433ff 17:11:12 sota_crawler v1 combo: burst+QAT before EMA + 15 GPTQ percentiles
  199. 12b55384 17:11:12 sota_crawler v1 combo: burst+QAT before EMA + 15 GPTQ percentiles
  200. 0ee493d2 17:11:12 sota_crawler v1 combo: burst+QAT before EMA + 15 GPTQ percentiles
  201. ab6c889f 17:17:22 parameter-golf-lab Record: 11L TTT Burst + GPTQ-15 + EMA + QAT (val_bpb=1.1232, seed 1337)
  202. ab6c889f 17:17:22 sota_crawler Record: 11L TTT Burst + GPTQ-15 + EMA + QAT (val_bpb=1.1232, seed 1337)
  203. 481f3501 17:20:07 parameter-golf-lab Isolate TTT to fractal cadence test cadence test layers only: blocks + loop_pos + skip_weights
  204. e239db2f 17:20:07 parameter-golf-lab Isolate TTT to fractal cadence test cadence test layers only: blocks + loop_pos + skip_weights
  205. f45beefc 17:20:07 parameter-golf-lab Isolate TTT to fractal cadence test cadence test layers only: blocks + loop_pos + skip_weights
  206. 7236a59d 17:20:07 parameter-golf-lab Isolate TTT to fractal cadence test cadence test layers only: blocks + loop_pos + skip_weights
  207. 481f3501 17:20:07 sota_crawler Isolate TTT to fractal cadence test cadence test layers only: blocks + loop_pos + skip_weights
  208. f45beefc 17:20:07 sota_crawler Isolate TTT to fractal cadence test cadence test layers only: blocks + loop_pos + skip_weights
  209. 7236a59d 17:20:07 sota_crawler Isolate TTT to fractal cadence test cadence test layers only: blocks + loop_pos + skip_weights
  210. 62f8f758 17:27:04 parameter-golf-lab Add TTT drift gate: snap block weights back toward trained originals
  211. b5c90dd2 17:27:04 parameter-golf-lab Add TTT drift gate: snap block weights back toward trained originals
  212. 66ebae7d 17:27:04 parameter-golf-lab Add TTT drift gate: snap block weights back toward trained originals
  213. 0e6fd680 17:27:04 parameter-golf-lab Add TTT drift gate: snap block weights back toward trained originals
  214. 62f8f758 17:27:04 sota_crawler Add TTT drift gate: snap block weights back toward trained originals
  215. 66ebae7d 17:27:04 sota_crawler Add TTT drift gate: snap block weights back toward trained originals
  216. 0e6fd680 17:27:04 sota_crawler Add TTT drift gate: snap block weights back toward trained originals
  217. 131166a9 17:32:37 parameter-golf-lab Update submission with accurate 2-seed results and correct train_gpt.py
  218. 131166a9 17:32:37 sota_crawler Update submission with accurate 2-seed results and correct train_gpt.py
  219. d9eb0497 17:32:42 parameter-golf-lab Remove accidentally staged files from submission branch
  220. d9eb0497 17:32:42 sota_crawler Remove accidentally staged files from submission branch
  221. ad9f7bea 17:48:43 parameter-golf-lab Add v2 self-distillation variant (val_bpb=1.1233, seed 1337)
  222. ad9f7bea 17:48:43 sota_crawler Add v2 self-distillation variant (val_bpb=1.1233, seed 1337)
  223. 75bb432d 17:52:47 parameter-golf-lab Add v4: burst + distill + train_seq_len=1024 for more steps
  224. 3c6f5bc1 17:52:47 parameter-golf-lab Add v4: burst + distill + train_seq_len=1024 for more steps
  225. 39c346b1 17:52:47 parameter-golf-lab Add v4: burst + distill + train_seq_len=1024 for more steps
  226. 50638c64 17:52:47 parameter-golf-lab Add v4: burst + distill + train_seq_len=1024 for more steps
  227. 75bb432d 17:52:47 sota_crawler Add v4: burst + distill + train_seq_len=1024 for more steps
  228. 39c346b1 17:52:47 sota_crawler Add v4: burst + distill + train_seq_len=1024 for more steps
  229. 50638c64 17:52:47 sota_crawler Add v4: burst + distill + train_seq_len=1024 for more steps
  230. 9cad0e01 18:01:23 parameter-golf-lab Add SOTA autoresearch: Qwen-guided edge finding for 1.1233 record
  231. c75145c6 18:01:23 parameter-golf-lab Add SOTA autoresearch: Qwen-guided edge finding for 1.1233 record
  232. f2ed46ee 18:01:23 parameter-golf-lab Add SOTA autoresearch: Qwen-guided edge finding for 1.1233 record
  233. 1529671e 18:01:23 parameter-golf-lab Add SOTA autoresearch: Qwen-guided edge finding for 1.1233 record
  234. 9cad0e01 18:01:23 sota_crawler Add SOTA autoresearch: Qwen-guided edge finding for 1.1233 record
  235. f2ed46ee 18:01:23 sota_crawler Add SOTA autoresearch: Qwen-guided edge finding for 1.1233 record
  236. 1529671e 18:01:23 sota_crawler Add SOTA autoresearch: Qwen-guided edge finding for 1.1233 record
  237. 39ad4e9a 18:05:27 parameter-golf-lab Fix TTT graph bug + upgrade to 3 blocks x 896d (23M params, ~14MB)
  238. 5808dddd 18:05:27 parameter-golf-lab Fix TTT graph bug + upgrade to 3 blocks x 896d (23M params, ~14MB)
  239. 2ba8b6fd 18:05:27 parameter-golf-lab Fix TTT graph bug + upgrade to 3 blocks x 896d (23M params, ~14MB)
  240. 1f65df52 18:05:27 parameter-golf-lab Fix TTT graph bug + upgrade to 3 blocks x 896d (23M params, ~14MB)
  241. 39ad4e9a 18:05:27 sota_crawler Fix TTT graph bug + upgrade to 3 blocks x 896d (23M params, ~14MB)
  242. 2ba8b6fd 18:05:27 sota_crawler Fix TTT graph bug + upgrade to 3 blocks x 896d (23M params, ~14MB)
  243. 1f65df52 18:05:27 sota_crawler Fix TTT graph bug + upgrade to 3 blocks x 896d (23M params, ~14MB)
  244. af6b29a5 18:22:48 parameter-golf-lab Save leapfrog experiment results: 6 variants tested, v1 wins at 1.12319
  245. 1a9b6816 18:22:48 parameter-golf-lab Save leapfrog experiment results: 6 variants tested, v1 wins at 1.12319
  246. c3b0643e 18:22:48 parameter-golf-lab Save leapfrog experiment results: 6 variants tested, v1 wins at 1.12319
  247. 041a8543 18:22:48 parameter-golf-lab Save leapfrog experiment results: 6 variants tested, v1 wins at 1.12319
  248. af6b29a5 18:22:48 sota_crawler Save leapfrog experiment results: 6 variants tested, v1 wins at 1.12319
  249. c3b0643e 18:22:48 sota_crawler Save leapfrog experiment results: 6 variants tested, v1 wins at 1.12319
  250. 041a8543 18:22:48 sota_crawler Save leapfrog experiment results: 6 variants tested, v1 wins at 1.12319
  251. fbfa775d 18:36:51 parameter-golf-lab fractal cadence test cadence test v2: dim=960 (25.7M, ~15.8MB), Muon LR back to 0.025
  252. 3bb985c6 18:36:51 parameter-golf-lab fractal cadence test cadence test v2: dim=960 (25.7M, ~15.8MB), Muon LR back to 0.025
  253. c3eda014 18:36:51 parameter-golf-lab fractal cadence test cadence test v2: dim=960 (25.7M, ~15.8MB), Muon LR back to 0.025
  254. d14df67c 18:36:51 parameter-golf-lab fractal cadence test cadence test v2: dim=960 (25.7M, ~15.8MB), Muon LR back to 0.025
  255. fbfa775d 18:36:51 sota_crawler fractal cadence test cadence test v2: dim=960 (25.7M, ~15.8MB), Muon LR back to 0.025
  256. c3eda014 18:36:51 sota_crawler fractal cadence test cadence test v2: dim=960 (25.7M, ~15.8MB), Muon LR back to 0.025
  257. d14df67c 18:36:51 sota_crawler fractal cadence test cadence test v2: dim=960 (25.7M, ~15.8MB), Muon LR back to 0.025
  258. cc78f04b 18:39:26 parameter-golf-lab Disable TTT in v2 (backward graph bug), focus on bigger model
  259. 41d743f5 18:39:26 parameter-golf-lab Disable TTT in v2 (backward graph bug), focus on bigger model
  260. 1a06df24 18:39:26 parameter-golf-lab Disable TTT in v2 (backward graph bug), focus on bigger model
  261. 7b5232a5 18:39:26 parameter-golf-lab Disable TTT in v2 (backward graph bug), focus on bigger model
  262. cc78f04b 18:39:26 sota_crawler Disable TTT in v2 (backward graph bug), focus on bigger model
  263. 1a06df24 18:39:26 sota_crawler Disable TTT in v2 (backward graph bug), focus on bigger model
  264. 7b5232a5 18:39:26 sota_crawler Disable TTT in v2 (backward graph bug), focus on bigger model
  265. 7d9b1d9c 18:40:17 parameter-golf-lab Add v5: QAT percentile fix + TrigramHash + EMA-SWA blend
  266. 6a6802b9 18:40:17 parameter-golf-lab Add v5: QAT percentile fix + TrigramHash + EMA-SWA blend
  267. e5834b45 18:40:17 parameter-golf-lab Add v5: QAT percentile fix + TrigramHash + EMA-SWA blend
  268. 3699d190 18:40:17 parameter-golf-lab Add v5: QAT percentile fix + TrigramHash + EMA-SWA blend
  269. 7d9b1d9c 18:40:17 sota_crawler Add v5: QAT percentile fix + TrigramHash + EMA-SWA blend
  270. e5834b45 18:40:17 sota_crawler Add v5: QAT percentile fix + TrigramHash + EMA-SWA blend
  271. 3699d190 18:40:17 sota_crawler Add v5: QAT percentile fix + TrigramHash + EMA-SWA blend
  272. 9c9a2e1e 18:41:23 parameter-golf-lab Fix TTT graph bug: fresh vc + detached inputs per loop
  273. 7e1220db 18:41:23 parameter-golf-lab Fix TTT graph bug: fresh vc + detached inputs per loop
  274. 9273a987 18:41:23 parameter-golf-lab Fix TTT graph bug: fresh vc + detached inputs per loop
  275. 79abf6cf 18:41:23 parameter-golf-lab Fix TTT graph bug: fresh vc + detached inputs per loop
  276. 9c9a2e1e 18:41:23 sota_crawler Fix TTT graph bug: fresh vc + detached inputs per loop
  277. 9273a987 18:41:23 sota_crawler Fix TTT graph bug: fresh vc + detached inputs per loop
  278. 79abf6cf 18:41:23 sota_crawler Fix TTT graph bug: fresh vc + detached inputs per loop
  279. 63c9b9e2 18:55:19 parameter-golf-lab Two v2 scripts: with TTT and without, for A/B comparison
  280. 05000048 18:55:19 parameter-golf-lab Two v2 scripts: with TTT and without, for A/B comparison
  281. b73d2298 18:55:19 parameter-golf-lab Two v2 scripts: with TTT and without, for A/B comparison
  282. bf17d79b 18:55:19 parameter-golf-lab Two v2 scripts: with TTT and without, for A/B comparison
  283. 63c9b9e2 18:55:19 sota_crawler Two v2 scripts: with TTT and without, for A/B comparison
  284. b73d2298 18:55:19 sota_crawler Two v2 scripts: with TTT and without, for A/B comparison
  285. bf17d79b 18:55:19 sota_crawler Two v2 scripts: with TTT and without, for A/B comparison
  286. b4a02886 19:14:04 parameter-golf-lab fractal cadence test cadence test v3: MLP 3.0->3.3 fills 16MB budget (27.4M params, ~15.2MB)
  287. 0243779c 19:14:04 parameter-golf-lab fractal cadence test cadence test v3: MLP 3.0->3.3 fills 16MB budget (27.4M params, ~15.2MB)
  288. 52c49dc5 19:14:04 parameter-golf-lab fractal cadence test cadence test v3: MLP 3.0->3.3 fills 16MB budget (27.4M params, ~15.2MB)
  289. 61aee008 19:14:04 parameter-golf-lab fractal cadence test cadence test v3: MLP 3.0->3.3 fills 16MB budget (27.4M params, ~15.2MB)
  290. b4a02886 19:14:04 sota_crawler fractal cadence test cadence test v3: MLP 3.0->3.3 fills 16MB budget (27.4M params, ~15.2MB)
  291. 52c49dc5 19:14:04 sota_crawler fractal cadence test cadence test v3: MLP 3.0->3.3 fills 16MB budget (27.4M params, ~15.2MB)
  292. 61aee008 19:14:04 sota_crawler fractal cadence test cadence test v3: MLP 3.0->3.3 fills 16MB budget (27.4M params, ~15.2MB)
  293. a23d65ae 19:32:33 parameter-golf-lab fractal cadence test cadence test v4: 1.5x batch (1.18M tokens/step) to use spare VRAM
  294. a7348276 19:32:33 parameter-golf-lab fractal cadence test cadence test v4: 1.5x batch (1.18M tokens/step) to use spare VRAM
  295. c7dac39c 19:32:33 parameter-golf-lab fractal cadence test cadence test v4: 1.5x batch (1.18M tokens/step) to use spare VRAM
  296. 672d0375 19:32:33 parameter-golf-lab fractal cadence test cadence test v4: 1.5x batch (1.18M tokens/step) to use spare VRAM
  297. a23d65ae 19:32:33 sota_crawler fractal cadence test cadence test v4: 1.5x batch (1.18M tokens/step) to use spare VRAM
  298. c7dac39c 19:32:33 sota_crawler fractal cadence test cadence test v4: 1.5x batch (1.18M tokens/step) to use spare VRAM
  299. 672d0375 19:32:33 sota_crawler fractal cadence test cadence test v4: 1.5x batch (1.18M tokens/step) to use spare VRAM
  300. e7f7a502 19:35:14 parameter-golf-lab Log recursive weight-sharing transformer: fractal cadence test cadence test cadence baseline at 1.2113 BPB
  301. b42deae8 19:35:14 parameter-golf-lab Log recursive weight-sharing transformer: fractal cadence test cadence test cadence baseline at 1.2113 BPB
  302. c28434ed 19:35:14 parameter-golf-lab Log recursive weight-sharing transformer: fractal cadence test cadence test cadence baseline at 1.2113 BPB
  303. 5f9b9e9b 19:35:14 parameter-golf-lab Log recursive weight-sharing transformer: fractal cadence test cadence test cadence baseline at 1.2113 BPB
  304. e7f7a502 19:35:14 sota_crawler Log recursive weight-sharing transformer: fractal cadence test cadence test cadence baseline at 1.2113 BPB
  305. c28434ed 19:35:14 sota_crawler Log recursive weight-sharing transformer: fractal cadence test cadence test cadence baseline at 1.2113 BPB
  306. 5f9b9e9b 19:35:14 sota_crawler Log recursive weight-sharing transformer: fractal cadence test cadence test cadence baseline at 1.2113 BPB
  307. d435755c 19:51:21 parameter-golf-lab Add v6: fractal cadence test cadence test 6L×2 loops (12 effective) + 480d/16H/4xMLP
  308. 659707bc 19:51:21 parameter-golf-lab Add v6: fractal cadence test cadence test 6L×2 loops (12 effective) + 480d/16H/4xMLP
  309. 8a50d685 19:51:21 parameter-golf-lab Add v6: fractal cadence test cadence test 6L×2 loops (12 effective) + 480d/16H/4xMLP
  310. e7585446 19:51:21 parameter-golf-lab Add v6: fractal cadence test cadence test 6L×2 loops (12 effective) + 480d/16H/4xMLP
  311. d435755c 19:51:21 sota_crawler Add v6: fractal cadence test cadence test 6L×2 loops (12 effective) + 480d/16H/4xMLP
  312. 8a50d685 19:51:21 sota_crawler Add v6: fractal cadence test cadence test 6L×2 loops (12 effective) + 480d/16H/4xMLP
  313. e7585446 19:51:21 sota_crawler Add v6: fractal cadence test cadence test 6L×2 loops (12 effective) + 480d/16H/4xMLP
  314. 95053441 20:02:40 parameter-golf-lab recursive weight-sharing transformer v5: TTT warmup-then-freeze to capture 1.19 peak
  315. 9ea76e2e 20:02:40 parameter-golf-lab recursive weight-sharing transformer v5: TTT warmup-then-freeze to capture 1.19 peak
  316. feef78ee 20:02:40 parameter-golf-lab recursive weight-sharing transformer v5: TTT warmup-then-freeze to capture 1.19 peak
  317. f196e844 20:02:40 parameter-golf-lab recursive weight-sharing transformer v5: TTT warmup-then-freeze to capture 1.19 peak
  318. 95053441 20:02:40 sota_crawler recursive weight-sharing transformer v5: TTT warmup-then-freeze to capture 1.19 peak
  319. feef78ee 20:02:40 sota_crawler recursive weight-sharing transformer v5: TTT warmup-then-freeze to capture 1.19 peak
  320. f196e844 20:02:40 sota_crawler recursive weight-sharing transformer v5: TTT warmup-then-freeze to capture 1.19 peak
  321. 0a319512 20:03:10 parameter-golf-lab Log all recursive weight-sharing transformer results: v3 TTT peak 1.1901, v4 batch 1.2186
  322. bbbb285e 20:03:10 parameter-golf-lab Log all recursive weight-sharing transformer results: v3 TTT peak 1.1901, v4 batch 1.2186
  323. b60ee7ad 20:03:10 parameter-golf-lab Log all recursive weight-sharing transformer results: v3 TTT peak 1.1901, v4 batch 1.2186
  324. 7fba307b 20:03:10 parameter-golf-lab Log all recursive weight-sharing transformer results: v3 TTT peak 1.1901, v4 batch 1.2186
  325. 0a319512 20:03:10 sota_crawler Log all recursive weight-sharing transformer results: v3 TTT peak 1.1901, v4 batch 1.2186
  326. b60ee7ad 20:03:10 sota_crawler Log all recursive weight-sharing transformer results: v3 TTT peak 1.1901, v4 batch 1.2186
  327. 7fba307b 20:03:10 sota_crawler Log all recursive weight-sharing transformer results: v3 TTT peak 1.1901, v4 batch 1.2186
  328. 89890d91 20:04:27 parameter-golf-lab Fix v6: 512d/16H/8KV (head_dim=32, FA3 requires multiple of 8)
  329. 3e0182b7 20:04:27 parameter-golf-lab Fix v6: 512d/16H/8KV (head_dim=32, FA3 requires multiple of 8)
  330. 4d029957 20:04:27 parameter-golf-lab Fix v6: 512d/16H/8KV (head_dim=32, FA3 requires multiple of 8)
  331. 5f4eb5f4 20:04:27 parameter-golf-lab Fix v6: 512d/16H/8KV (head_dim=32, FA3 requires multiple of 8)
  332. 89890d91 20:04:27 sota_crawler Fix v6: 512d/16H/8KV (head_dim=32, FA3 requires multiple of 8)
  333. 4d029957 20:04:27 sota_crawler Fix v6: 512d/16H/8KV (head_dim=32, FA3 requires multiple of 8)
  334. 5f4eb5f4 20:04:27 sota_crawler Fix v6: 512d/16H/8KV (head_dim=32, FA3 requires multiple of 8)
  335. e26df9b8 20:24:20 parameter-golf-lab Log v5/v6 results + add fractal cadence test cadence test sweep script
  336. 1017b30b 20:24:20 parameter-golf-lab Log v5/v6 results + add fractal cadence test cadence test sweep script
  337. cb5acc1c 20:24:20 parameter-golf-lab Log v5/v6 results + add fractal cadence test cadence test sweep script
  338. 74eebb6a 20:24:20 parameter-golf-lab Log v5/v6 results + add fractal cadence test cadence test sweep script
  339. e26df9b8 20:24:20 sota_crawler Log v5/v6 results + add fractal cadence test cadence test sweep script
  340. cb5acc1c 20:24:20 sota_crawler Log v5/v6 results + add fractal cadence test cadence test sweep script
  341. 74eebb6a 20:24:20 sota_crawler Log v5/v6 results + add fractal cadence test cadence test sweep script
  342. df42c95c 20:56:11 parameter-golf-lab Add seed 2024: 1.1239 BPB, 15.50 MB — 3-seed validation complete
  343. df42c95c 20:56:11 sota_crawler Add seed 2024: 1.1239 BPB, 15.50 MB — 3-seed validation complete
  344. cbcd7f71 20:56:19 parameter-golf-lab Remove accidentally staged files
  345. cbcd7f71 20:56:19 sota_crawler Remove accidentally staged files
  346. a474a327 20:58:10 parameter-golf-lab Add recursive weight-sharing transformer deep calibration autoresearch
  347. 6b333a81 21:03:24 parameter-golf-lab recursive weight-sharing transformer 8-hour single GPU longrun
  348. 67ca84b8 21:03:24 parameter-golf-lab recursive weight-sharing transformer 8-hour single GPU longrun
  349. 26bb918f 21:03:24 parameter-golf-lab recursive weight-sharing transformer 8-hour single GPU longrun
  350. bd37700e 21:03:24 parameter-golf-lab recursive weight-sharing transformer 8-hour single GPU longrun
  351. 6b333a81 21:03:24 sota_crawler recursive weight-sharing transformer 8-hour single GPU longrun
  352. 26bb918f 21:03:24 sota_crawler recursive weight-sharing transformer 8-hour single GPU longrun
  353. bd37700e 21:03:24 sota_crawler recursive weight-sharing transformer 8-hour single GPU longrun
  354. aec30f3e 21:29:00 parameter-golf-lab Fix PolyAttention: use cubic x^3 with learned dynamic scaling per paper
  355. 0cba9336 22:19:31 parameter-golf-lab DGAttention v2: asymmetric D, EMA baseline, hybrid payload
  356. 91a95e78 22:25:43 parameter-golf-lab DG v2.1: replace sequential EMA loop with parallel cumsum baseline
  357. 475acf31 00:54:14 parameter-golf-lab Complete recursive weight-sharing transformer documentation: all runs, findings, architecture analysis
  358. e0527c1e 00:54:14 parameter-golf-lab Complete recursive weight-sharing transformer documentation: all runs, findings, architecture analysis
  359. 4394e446 00:54:14 parameter-golf-lab Complete recursive weight-sharing transformer documentation: all runs, findings, architecture analysis
  360. cc2ff3a3 00:54:14 parameter-golf-lab Complete recursive weight-sharing transformer documentation: all runs, findings, architecture analysis
  361. 475acf31 00:54:14 sota_crawler Complete recursive weight-sharing transformer documentation: all runs, findings, architecture analysis
  362. 4394e446 00:54:14 sota_crawler Complete recursive weight-sharing transformer documentation: all runs, findings, architecture analysis
  363. cc2ff3a3 00:54:14 sota_crawler Complete recursive weight-sharing transformer documentation: all runs, findings, architecture analysis
  364. 670ca139 00:55:00 parameter-golf-lab Add v7: legal score-first TTT eval (PR #461 recipe)
  365. b34f4249 00:55:00 parameter-golf-lab Add v7: legal score-first TTT eval (PR #461 recipe)
  366. 16336852 00:55:00 parameter-golf-lab Add v7: legal score-first TTT eval (PR #461 recipe)
  367. dec45942 00:55:00 parameter-golf-lab Add v7: legal score-first TTT eval (PR #461 recipe)
  368. 670ca139 00:55:00 sota_crawler Add v7: legal score-first TTT eval (PR #461 recipe)
  369. 16336852 00:55:00 sota_crawler Add v7: legal score-first TTT eval (PR #461 recipe)
  370. dec45942 00:55:00 sota_crawler Add v7: legal score-first TTT eval (PR #461 recipe)
  371. 8027566a 01:56:17 parameter-golf-lab Add v8: mutual learning — flat GPT + fractal cadence test cadence test GPT co-training
  372. 03849f53 01:56:17 parameter-golf-lab Add v8: mutual learning — flat GPT + fractal cadence test cadence test GPT co-training
  373. d03b8aad 01:56:17 parameter-golf-lab Add v8: mutual learning — flat GPT + fractal cadence test cadence test GPT co-training
  374. 735b22f7 01:56:17 parameter-golf-lab Add v8: mutual learning — flat GPT + fractal cadence test cadence test GPT co-training
  375. 8027566a 01:56:17 sota_crawler Add v8: mutual learning — flat GPT + fractal cadence test cadence test GPT co-training
  376. d03b8aad 01:56:17 sota_crawler Add v8: mutual learning — flat GPT + fractal cadence test cadence test GPT co-training
  377. 735b22f7 01:56:17 sota_crawler Add v8: mutual learning — flat GPT + fractal cadence test cadence test GPT co-training
  378. 0b97cd81 01:59:40 parameter-golf-lab v7: add TTT early stopping — train only first 60 chunks
  379. c4463514 01:59:40 parameter-golf-lab v7: add TTT early stopping — train only first 60 chunks
  380. 7e42a13e 01:59:40 parameter-golf-lab v7: add TTT early stopping — train only first 60 chunks
  381. 4fbad042 01:59:40 parameter-golf-lab v7: add TTT early stopping — train only first 60 chunks
  382. 0b97cd81 01:59:40 sota_crawler v7: add TTT early stopping — train only first 60 chunks
  383. 7e42a13e 01:59:40 sota_crawler v7: add TTT early stopping — train only first 60 chunks
  384. 4fbad042 01:59:40 sota_crawler v7: add TTT early stopping — train only first 60 chunks
  385. b23d437e 02:10:02 parameter-golf-lab Update v6 to sweep winner: 4×3/512d/8H/4KV/4xMLP
  386. 8fd43260 02:10:02 parameter-golf-lab Update v6 to sweep winner: 4×3/512d/8H/4KV/4xMLP
  387. 4ead1eed 02:10:02 parameter-golf-lab Update v6 to sweep winner: 4×3/512d/8H/4KV/4xMLP
  388. 6d68de38 02:10:02 parameter-golf-lab Update v6 to sweep winner: 4×3/512d/8H/4KV/4xMLP
  389. b23d437e 02:10:02 sota_crawler Update v6 to sweep winner: 4×3/512d/8H/4KV/4xMLP
  390. 4ead1eed 02:10:02 sota_crawler Update v6 to sweep winner: 4×3/512d/8H/4KV/4xMLP
  391. 6d68de38 02:10:02 sota_crawler Update v6 to sweep winner: 4×3/512d/8H/4KV/4xMLP
  392. b452f849 02:40:55 parameter-golf-lab Log v7 TTT results: early stop at 60 chunks = 1.12312 (best)
  393. 7c9d8cc8 02:40:55 parameter-golf-lab Log v7 TTT results: early stop at 60 chunks = 1.12312 (best)
  394. d3b1e2fd 02:40:55 parameter-golf-lab Log v7 TTT results: early stop at 60 chunks = 1.12312 (best)
  395. bd1e51d7 02:40:55 parameter-golf-lab Log v7 TTT results: early stop at 60 chunks = 1.12312 (best)
  396. b452f849 02:40:55 sota_crawler Log v7 TTT results: early stop at 60 chunks = 1.12312 (best)
  397. d3b1e2fd 02:40:55 sota_crawler Log v7 TTT results: early stop at 60 chunks = 1.12312 (best)
  398. bd1e51d7 02:40:55 sota_crawler Log v7 TTT results: early stop at 60 chunks = 1.12312 (best)
  399. 9fea2b6d 03:04:30 parameter-golf-lab The recursive weight-sharing depth variant: 6L x 2 loops, dim=640, MLP 4x, 28.5M params
  400. 19acd25d 03:04:30 parameter-golf-lab The recursive weight-sharing depth variant: 6L x 2 loops, dim=640, MLP 4x, 28.5M params
  401. 961dbbd2 03:04:30 parameter-golf-lab The recursive weight-sharing depth variant: 6L x 2 loops, dim=640, MLP 4x, 28.5M params
  402. afba0768 03:04:30 parameter-golf-lab The recursive weight-sharing depth variant: 6L x 2 loops, dim=640, MLP 4x, 28.5M params
  403. 9fea2b6d 03:04:30 sota_crawler The recursive weight-sharing depth variant: 6L x 2 loops, dim=640, MLP 4x, 28.5M params
  404. 961dbbd2 03:04:30 sota_crawler The recursive weight-sharing depth variant: 6L x 2 loops, dim=640, MLP 4x, 28.5M params
  405. afba0768 03:04:30 sota_crawler The recursive weight-sharing depth variant: 6L x 2 loops, dim=640, MLP 4x, 28.5M params
  406. cd3d5bc6 03:10:11 parameter-golf-lab Log higher LR (0.030) + MTP results — both worse
  407. 504d3d99 03:10:11 parameter-golf-lab Log higher LR (0.030) + MTP results — both worse
  408. c51d394e 03:10:11 parameter-golf-lab Log higher LR (0.030) + MTP results — both worse
  409. 4fafae88 03:10:11 parameter-golf-lab Log higher LR (0.030) + MTP results — both worse
  410. cd3d5bc6 03:10:11 sota_crawler Log higher LR (0.030) + MTP results — both worse
  411. c51d394e 03:10:11 sota_crawler Log higher LR (0.030) + MTP results — both worse
  412. 4fafae88 03:10:11 sota_crawler Log higher LR (0.030) + MTP results — both worse
  413. c0db6ddd 03:11:00 parameter-golf-lab Fix DG v2.1 dk reshape: use num_kv_heads * d_head_dim for dk projection
  414. 924b77ca 03:28:39 parameter-golf-lab The recursive weight-sharing depth variant: 1.1478 BPB — 0.025 from world lead
  415. c8552f43 03:28:39 parameter-golf-lab The recursive weight-sharing depth variant: 1.1478 BPB — 0.025 from world lead
  416. 0263f445 03:28:39 parameter-golf-lab The recursive weight-sharing depth variant: 1.1478 BPB — 0.025 from world lead
  417. c40bfce6 03:28:39 parameter-golf-lab The recursive weight-sharing depth variant: 1.1478 BPB — 0.025 from world lead
  418. 924b77ca 03:28:39 sota_crawler The recursive weight-sharing depth variant: 1.1478 BPB — 0.025 from world lead
  419. 0263f445 03:28:39 sota_crawler The recursive weight-sharing depth variant: 1.1478 BPB — 0.025 from world lead
  420. c40bfce6 03:28:39 sota_crawler The recursive weight-sharing depth variant: 1.1478 BPB — 0.025 from world lead
  421. f5b78fb4 03:30:34 parameter-golf-lab Stack distillation onto recursive weight-sharing depth variant
  422. 7f556e65 03:30:34 parameter-golf-lab Stack distillation onto recursive weight-sharing depth variant
  423. d60df798 03:30:34 parameter-golf-lab Stack distillation onto recursive weight-sharing depth variant
  424. 57062a5c 03:30:34 parameter-golf-lab Stack distillation onto recursive weight-sharing depth variant
  425. f5b78fb4 03:30:34 sota_crawler Stack distillation onto recursive weight-sharing depth variant
  426. d60df798 03:30:34 sota_crawler Stack distillation onto recursive weight-sharing depth variant
  427. 57062a5c 03:30:34 sota_crawler Stack distillation onto recursive weight-sharing depth variant
  428. 1ea90b74 03:34:25 parameter-golf-lab Prep recursive weight-sharing depth variant PR submission (draft — needs image)
  429. f1ac766e 03:34:25 parameter-golf-lab Prep recursive weight-sharing depth variant PR submission (draft — needs image)
  430. 868accf4 03:34:25 parameter-golf-lab Prep recursive weight-sharing depth variant PR submission (draft — needs image)
  431. eef3124d 03:34:25 parameter-golf-lab Prep recursive weight-sharing depth variant PR submission (draft — needs image)
  432. 1ea90b74 03:34:25 sota_crawler Prep recursive weight-sharing depth variant PR submission (draft — needs image)
  433. 868accf4 03:34:25 sota_crawler Prep recursive weight-sharing depth variant PR submission (draft — needs image)
  434. eef3124d 03:34:25 sota_crawler Prep recursive weight-sharing depth variant PR submission (draft — needs image)
  435. 9e0b23a4 03:36:31 parameter-golf-lab Add train script to recursive weight-sharing transformer submission folder
  436. edea2e13 03:36:31 parameter-golf-lab Add train script to recursive weight-sharing transformer submission folder
  437. 132f4494 03:36:31 parameter-golf-lab Add train script to recursive weight-sharing transformer submission folder
  438. 2c918975 03:36:31 parameter-golf-lab Add train script to recursive weight-sharing transformer submission folder
  439. 9e0b23a4 03:36:31 sota_crawler Add train script to recursive weight-sharing transformer submission folder
  440. 132f4494 03:36:31 sota_crawler Add train script to recursive weight-sharing transformer submission folder
  441. 2c918975 03:36:31 sota_crawler Add train script to recursive weight-sharing transformer submission folder
  442. fb530e28 03:51:24 parameter-golf-lab TTT stabilization: EMA scoring + fixed cosine LR + embed freeze
  443. fe27bfb8 03:51:24 parameter-golf-lab TTT stabilization: EMA scoring + fixed cosine LR + embed freeze
  444. 5229c66a 03:51:24 parameter-golf-lab TTT stabilization: EMA scoring + fixed cosine LR + embed freeze
  445. cb4f8790 03:51:24 parameter-golf-lab TTT stabilization: EMA scoring + fixed cosine LR + embed freeze
  446. fb530e28 03:51:24 sota_crawler TTT stabilization: EMA scoring + fixed cosine LR + embed freeze
  447. 5229c66a 03:51:24 sota_crawler TTT stabilization: EMA scoring + fixed cosine LR + embed freeze
  448. cb4f8790 03:51:24 sota_crawler TTT stabilization: EMA scoring + fixed cosine LR + embed freeze
  449. 46894842 04:17:43 parameter-golf-lab recursive weight-sharing transformer: Recursive Weight Sharing + MLP 4x (1.1478 BPB)
  450. 37aa1fe4 04:17:43 parameter-golf-lab recursive weight-sharing transformer: Recursive Weight Sharing + MLP 4x (1.1478 BPB)
  451. 46894842 04:17:43 sota_crawler recursive weight-sharing transformer: Recursive Weight Sharing + MLP 4x (1.1478 BPB)
  452. 37aa1fe4 04:17:43 sota_crawler recursive weight-sharing transformer: Recursive Weight Sharing + MLP 4x (1.1478 BPB)
  453. 08635604 04:21:57 parameter-golf-lab Add train log to recursive weight-sharing transformer submission
  454. 5305a17c 04:21:57 parameter-golf-lab Add train log to recursive weight-sharing transformer submission
  455. 08635604 04:21:57 sota_crawler Add train log to recursive weight-sharing transformer submission
  456. 5305a17c 04:21:57 sota_crawler Add train log to recursive weight-sharing transformer submission
  457. 93205f17 04:23:45 parameter-golf-lab Add GPTQ quantization: Hessian-aware error compensation for int6
  458. fe467672 04:23:45 parameter-golf-lab Add GPTQ quantization: Hessian-aware error compensation for int6
  459. f20e3876 04:23:45 parameter-golf-lab Add GPTQ quantization: Hessian-aware error compensation for int6
  460. ffad368a 04:23:45 parameter-golf-lab Add GPTQ quantization: Hessian-aware error compensation for int6
  461. 93205f17 04:23:45 sota_crawler Add GPTQ quantization: Hessian-aware error compensation for int6
  462. f20e3876 04:23:45 sota_crawler Add GPTQ quantization: Hessian-aware error compensation for int6
  463. ffad368a 04:23:45 sota_crawler Add GPTQ quantization: Hessian-aware error compensation for int6
  464. 939c140c 04:29:03 parameter-golf-lab Stack low-hanging fruit: fix GPTQ, earlier QAT, matched clipping
  465. 5ef90aef 04:29:03 parameter-golf-lab Stack low-hanging fruit: fix GPTQ, earlier QAT, matched clipping
  466. 011e7d94 04:29:03 parameter-golf-lab Stack low-hanging fruit: fix GPTQ, earlier QAT, matched clipping
  467. 8e0a01f7 04:29:03 parameter-golf-lab Stack low-hanging fruit: fix GPTQ, earlier QAT, matched clipping
  468. 939c140c 04:29:03 sota_crawler Stack low-hanging fruit: fix GPTQ, earlier QAT, matched clipping
  469. 011e7d94 04:29:03 sota_crawler Stack low-hanging fruit: fix GPTQ, earlier QAT, matched clipping
  470. 8e0a01f7 04:29:03 sota_crawler Stack low-hanging fruit: fix GPTQ, earlier QAT, matched clipping
  471. e523ad13 04:32:18 parameter-golf-lab DG v2.2: blended baseline (previous token + global mean)
  472. 177e03c8 04:39:30 parameter-golf-lab Add per-layer β/γ diagnostic logging for DG attention
  473. e15e5edf 04:59:35 parameter-golf-lab Add post-quant training burst: repair quant damage before eval
  474. 17b34e91 04:59:35 parameter-golf-lab Add post-quant training burst: repair quant damage before eval
  475. 988e83b0 04:59:35 parameter-golf-lab Add post-quant training burst: repair quant damage before eval
  476. 4f3bde64 04:59:35 parameter-golf-lab Add post-quant training burst: repair quant damage before eval
  477. e15e5edf 04:59:35 sota_crawler Add post-quant training burst: repair quant damage before eval
  478. 988e83b0 04:59:35 sota_crawler Add post-quant training burst: repair quant damage before eval
  479. 4f3bde64 04:59:35 sota_crawler Add post-quant training burst: repair quant damage before eval
DATE2026-03-22
DAY005
PRIMARY METRICVAL_BPB
TESTS RUN15 unique (15 valid / 0 TTT leakage)
ARCHITECTURE FAMILYLEADERBOARD TRANSFORMER VARIANTS (×4) · TTT8 STRIDE-32 TEST FAMILY (×3) · RECURSIVE WEIGHT-SHARING TRANSFORMER (×2) · TTT-ONLY MODEL TEST (×1) · PROGRESSIVE MLP AND FRACTAL CADENCE TESTS (×1) · STATE-SPACE TTT VARIANT (×1) · +3 MORE (×3)
CLASSIFICATIONEXPERIMENTAL 80% (12)/NEURAL 20% (3)

SECTION 1.0 — TRAJECTORY

VAL_BPB for 2026-03-22 panel 1 of 2
VAL_BPB for 2026-03-22 panel 2 of 2
VAL_BPB SPREADnminmedianmaxdownload
LEADERBOARD TRANSFORMER VARIANTS41.12321.12331.1243
TTT8 STRIDE-32 TEST FAMILY31.13011.13071.1313download
RECURSIVE WEIGHT-SHARING TRANSFORMER21.14781.14781.1478download
TTT-ONLY MODEL TEST11.12401.12401.1240download
PROGRESSIVE MLP AND FRACTAL CADENCE TESTS11.65751.65751.6575
STATE-SPACE TTT VARIANT11.12231.12231.1223
POST-TRAINING QUANTIZATION CALIBRATION11.12321.12321.1232download
CRAWLER ABLATION CONTROL SET11.21131.21131.2113
OTHER11.12321.12321.1232
ALL VALID151.12231.12431.6575

SECTION 2.0 — TEST RUNS

EXPERIMENTAL TEST RUNS — 12

#TIMEHWABLATIONVAL_BPBSTATUSΔ %OPEN
100:488×H100run_s1337.log (×6)1.1240OK↑0.3%
training_log 1
training_log 2
training_log 3
training_log 4
training_log 5
training_log 6
201:458×H100run_s42.log (×2)1.1307OK↑0.9%
training_log 1
training_log 2
302:058×H100run_s7.log1.1313OK↑0.9%
training_log
402:228×H100run_s137.log1.1301OK↑0.8%
training_log
503:28The recursive weight-sharing depth variant: 1.1478 BPB — 0.025 from world lea…1.1478OK↑2.4%
details
604:17recursive weight-sharing transformer: Recursive Weight Sharing + MLP 4x (1.14…1.1478OK↑2.4%
details
706:04Add fractal cadence test cadence test cadence long run (1.6575 BPB @ 3929 ste…1.6575OK↑47.9%
details
806:55Log depth + TTT results: 1.1223 BPB (12L+TTT20ep+SAM) (×7)1.1223OK↑0.1%
details
1016:19Add 3 leapfrog variants based on PR#414 SOTA (1.1233 BPB) (×7)1.1233OK↑0.2%
details
1317:48Add v2 self-distillation variant (val_bpb=1.1233, seed 1337) (×2)1.1233OK↑0.2%
details
1419:35Log recursive weight-sharing transformer: fractal cadence test cadence test c…1.2113OK↑8.1%
details
1520:56Add seed 2024: 1.1239 BPB, 15.50 MB — 3-seed validation complete (×2)1.1232OK↑0.2%
details

NEURAL TEST RUNS — 3

#TIMEHWABLATIONVAL_BPBSTATUSΔ %OPEN
907:09Save pr374_throne: our #1 non-TTT record holder (1.1243 BPB) (×7)1.1243OK↑0.3%
details
1117:17Record: 11L TTT Burst + GPTQ-15 + EMA + QAT (val_bpb=1.1232, seed 1337) (×2)1.1232OK↑0.2%
details
1217:32Update submission with accurate 2-seed results and correct train_gpt.py (×2)1.1232OK↑0.2%
details