DAILY RESEARCH REPORTBACK TO MASTER SHEET2026-03-23

val_bpb · 568 of 658 · 12 warnings
| 2026-03-23 | 12 | 22 |
{
  "date": "2026-03-23",
  "fatal": 0,
  "headline_metric": "val_bpb",
  "raw_records": 658,
  "render_records": 568,
  "test_rows": 22,
  "warnings": 12
}
256 commits on 2026-03-23
  1. 899a1db8 05:06:31 parameter-golf-lab Remove post-quant burst — illegal (uses training data during eval)
  2. 711eb3a6 05:06:31 parameter-golf-lab Remove post-quant burst — illegal (uses training data during eval)
  3. fcfbd34f 05:06:31 parameter-golf-lab Remove post-quant burst — illegal (uses training data during eval)
  4. 217eaad5 05:06:31 parameter-golf-lab Remove post-quant burst — illegal (uses training data during eval)
  5. 899a1db8 05:06:31 sota_crawler Remove post-quant burst — illegal (uses training data during eval)
  6. fcfbd34f 05:06:31 sota_crawler Remove post-quant burst — illegal (uses training data during eval)
  7. 217eaad5 05:06:31 sota_crawler Remove post-quant burst — illegal (uses training data during eval)
  8. 3c049dde 05:14:56 parameter-golf-lab Add selective int8 for sensitive layers (GPTQ-int8 upgrade path)
  9. 0a22c2b7 05:14:56 parameter-golf-lab Add selective int8 for sensitive layers (GPTQ-int8 upgrade path)
  10. 286b2adc 05:14:56 parameter-golf-lab Add selective int8 for sensitive layers (GPTQ-int8 upgrade path)
  11. f94c09a3 05:14:56 parameter-golf-lab Add selective int8 for sensitive layers (GPTQ-int8 upgrade path)
  12. 3c049dde 05:14:56 sota_crawler Add selective int8 for sensitive layers (GPTQ-int8 upgrade path)
  13. 286b2adc 05:14:56 sota_crawler Add selective int8 for sensitive layers (GPTQ-int8 upgrade path)
  14. f94c09a3 05:14:56 sota_crawler Add selective int8 for sensitive layers (GPTQ-int8 upgrade path)
  15. fc900eb2 05:18:39 parameter-golf-lab recursive weight-sharing transformer V3: v7 full stack + fractal cadence test cadence test loops (GPTQ + TTT + PQB)
  16. 769b01b5 05:18:39 parameter-golf-lab recursive weight-sharing transformer V3: v7 full stack + fractal cadence test cadence test loops (GPTQ + TTT + PQB)
  17. c96450e2 05:18:39 parameter-golf-lab recursive weight-sharing transformer V3: v7 full stack + fractal cadence test cadence test loops (GPTQ + TTT + PQB)
  18. 8b22b10f 05:18:39 parameter-golf-lab recursive weight-sharing transformer V3: v7 full stack + fractal cadence test cadence test loops (GPTQ + TTT + PQB)
  19. fc900eb2 05:18:39 sota_crawler recursive weight-sharing transformer V3: v7 full stack + fractal cadence test cadence test loops (GPTQ + TTT + PQB)
  20. c96450e2 05:18:39 sota_crawler recursive weight-sharing transformer V3: v7 full stack + fractal cadence test cadence test loops (GPTQ + TTT + PQB)
  21. 8b22b10f 05:18:39 sota_crawler recursive weight-sharing transformer V3: v7 full stack + fractal cadence test cadence test loops (GPTQ + TTT + PQB)
  22. 6d4772ce 05:54:45 parameter-golf-lab Add AdamW TTT option — PR #462 shows 5x better TTT gain vs SGD
  23. b5986e4d 05:54:45 parameter-golf-lab Add AdamW TTT option — PR #462 shows 5x better TTT gain vs SGD
  24. ff4a79ef 05:54:45 parameter-golf-lab Add AdamW TTT option — PR #462 shows 5x better TTT gain vs SGD
  25. a0e0841e 05:54:45 parameter-golf-lab Add AdamW TTT option — PR #462 shows 5x better TTT gain vs SGD
  26. 6d4772ce 05:54:45 sota_crawler Add AdamW TTT option — PR #462 shows 5x better TTT gain vs SGD
  27. ff4a79ef 05:54:45 sota_crawler Add AdamW TTT option — PR #462 shows 5x better TTT gain vs SGD
  28. a0e0841e 05:54:45 sota_crawler Add AdamW TTT option — PR #462 shows 5x better TTT gain vs SGD
  29. e605d1c5 06:02:05 parameter-golf-lab Record: GPTQ + Early QAT + TTT EMA — 3-seed mean val_bpb 1.1215
  30. 464f8591 06:02:05 parameter-golf-lab Record: GPTQ + Early QAT + TTT EMA — 3-seed mean val_bpb 1.1215
  31. 464f8591 06:02:05 sota_crawler Record: GPTQ + Early QAT + TTT EMA — 3-seed mean val_bpb 1.1215
  32. cd7871ef 06:04:46 parameter-golf-lab recursive weight-sharing transformer V3: SwiGLU + AdamW TTT + 8K bigram (PR#462 techniques)
  33. 8fe38f0a 06:04:46 parameter-golf-lab recursive weight-sharing transformer V3: SwiGLU + AdamW TTT + 8K bigram (PR#462 techniques)
  34. 8a195607 06:04:46 parameter-golf-lab recursive weight-sharing transformer V3: SwiGLU + AdamW TTT + 8K bigram (PR#462 techniques)
  35. cd7871ef 06:04:46 sota_crawler recursive weight-sharing transformer V3: SwiGLU + AdamW TTT + 8K bigram (PR#462 techniques)
  36. 8fe38f0a 06:04:46 sota_crawler recursive weight-sharing transformer V3: SwiGLU + AdamW TTT + 8K bigram (PR#462 techniques)
  37. 8a195607 06:04:46 sota_crawler recursive weight-sharing transformer V3: SwiGLU + AdamW TTT + 8K bigram (PR#462 techniques)
  38. 1f6d92a7 06:21:26 parameter-golf-lab Fix DDP unused params: add find_unused_parameters=True
  39. 98f3e2f5 06:21:26 parameter-golf-lab Fix DDP unused params: add find_unused_parameters=True
  40. a5724c0d 06:21:26 parameter-golf-lab Fix DDP unused params: add find_unused_parameters=True
  41. 1f6d92a7 06:21:26 sota_crawler Fix DDP unused params: add find_unused_parameters=True
  42. 98f3e2f5 06:21:26 sota_crawler Fix DDP unused params: add find_unused_parameters=True
  43. a5724c0d 06:21:26 sota_crawler Fix DDP unused params: add find_unused_parameters=True
  44. c4763c7d 06:26:42 parameter-golf-lab SwiGLU + GPTQ: PR #462 architecture with our quant improvements
  45. c97ad546 06:26:42 parameter-golf-lab SwiGLU + GPTQ: PR #462 architecture with our quant improvements
  46. 04eb2e54 06:26:42 parameter-golf-lab SwiGLU + GPTQ: PR #462 architecture with our quant improvements
  47. c4763c7d 06:26:42 sota_crawler SwiGLU + GPTQ: PR #462 architecture with our quant improvements
  48. c97ad546 06:26:42 sota_crawler SwiGLU + GPTQ: PR #462 architecture with our quant improvements
  49. 04eb2e54 06:26:42 sota_crawler SwiGLU + GPTQ: PR #462 architecture with our quant improvements
  50. dbd79b65 06:43:47 parameter-golf-lab Add OptRot: Hadamard rotation before GPTQ quantization
  51. 03e5f541 06:43:47 parameter-golf-lab Add OptRot: Hadamard rotation before GPTQ quantization
  52. b639e15b 06:43:47 parameter-golf-lab Add OptRot: Hadamard rotation before GPTQ quantization
  53. dbd79b65 06:43:47 sota_crawler Add OptRot: Hadamard rotation before GPTQ quantization
  54. 03e5f541 06:43:47 sota_crawler Add OptRot: Hadamard rotation before GPTQ quantization
  55. b639e15b 06:43:47 sota_crawler Add OptRot: Hadamard rotation before GPTQ quantization
  56. 675f3749 06:49:47 parameter-golf-lab recursive weight-sharing transformer V3: Star-ReLU SwiGLU + VRL + MLP 4.0 (fits 16MB)
  57. eb09f7ec 06:49:47 parameter-golf-lab recursive weight-sharing transformer V3: Star-ReLU SwiGLU + VRL + MLP 4.0 (fits 16MB)
  58. 5aec5ae3 06:49:47 parameter-golf-lab recursive weight-sharing transformer V3: Star-ReLU SwiGLU + VRL + MLP 4.0 (fits 16MB)
  59. 675f3749 06:49:47 sota_crawler recursive weight-sharing transformer V3: Star-ReLU SwiGLU + VRL + MLP 4.0 (fits 16MB)
  60. eb09f7ec 06:49:47 sota_crawler recursive weight-sharing transformer V3: Star-ReLU SwiGLU + VRL + MLP 4.0 (fits 16MB)
  61. 5aec5ae3 06:49:47 sota_crawler recursive weight-sharing transformer V3: Star-ReLU SwiGLU + VRL + MLP 4.0 (fits 16MB)
  62. 34335573 07:12:44 parameter-golf-lab Log session results: 1.0763 SwiGLU (size problem), 1.1215 v7 (submitted)
  63. 7020bf6d 07:12:44 parameter-golf-lab Log session results: 1.0763 SwiGLU (size problem), 1.1215 v7 (submitted)
  64. ce487d52 07:12:44 parameter-golf-lab Log session results: 1.0763 SwiGLU (size problem), 1.1215 v7 (submitted)
  65. 34335573 07:12:44 sota_crawler Log session results: 1.0763 SwiGLU (size problem), 1.1215 v7 (submitted)
  66. 7020bf6d 07:12:44 sota_crawler Log session results: 1.0763 SwiGLU (size problem), 1.1215 v7 (submitted)
  67. ce487d52 07:12:44 sota_crawler Log session results: 1.0763 SwiGLU (size problem), 1.1215 v7 (submitted)
  68. 60137e91 13:39:04 parameter-golf-lab recursive weight-sharing transformer V4: GEPA-matched config for 16MB compression
  69. addbea95 13:39:04 parameter-golf-lab recursive weight-sharing transformer V4: GEPA-matched config for 16MB compression
  70. 06952977 13:39:04 parameter-golf-lab recursive weight-sharing transformer V4: GEPA-matched config for 16MB compression
  71. 60137e91 13:39:04 sota_crawler recursive weight-sharing transformer V4: GEPA-matched config for 16MB compression
  72. addbea95 13:39:04 sota_crawler recursive weight-sharing transformer V4: GEPA-matched config for 16MB compression
  73. 06952977 13:39:04 sota_crawler recursive weight-sharing transformer V4: GEPA-matched config for 16MB compression
  74. e90f9c16 13:54:29 parameter-golf-lab H100-ready: revert to DG v2.1, auto-detect GQA, clean up
  75. 01e7c574 13:57:54 parameter-golf-lab recursive weight-sharing transformer V4 A/B: 5x2 and 4x2 variants for H100 calibration
  76. 5c912b51 13:57:54 parameter-golf-lab recursive weight-sharing transformer V4 A/B: 5x2 and 4x2 variants for H100 calibration
  77. 5b72e979 13:57:54 parameter-golf-lab recursive weight-sharing transformer V4 A/B: 5x2 and 4x2 variants for H100 calibration
  78. 01e7c574 13:57:54 sota_crawler recursive weight-sharing transformer V4 A/B: 5x2 and 4x2 variants for H100 calibration
  79. 5c912b51 13:57:54 sota_crawler recursive weight-sharing transformer V4 A/B: 5x2 and 4x2 variants for H100 calibration
  80. 5b72e979 13:57:54 sota_crawler recursive weight-sharing transformer V4 A/B: 5x2 and 4x2 variants for H100 calibration
  81. 2c594466 14:06:12 parameter-golf-lab Add short TTT experiment: no-EMA, 50 chunks, SGD
  82. 324843ba 14:06:12 parameter-golf-lab Add short TTT experiment: no-EMA, 50 chunks, SGD
  83. eec0981a 14:06:12 parameter-golf-lab Add short TTT experiment: no-EMA, 50 chunks, SGD
  84. 2c594466 14:06:12 sota_crawler Add short TTT experiment: no-EMA, 50 chunks, SGD
  85. 324843ba 14:06:12 sota_crawler Add short TTT experiment: no-EMA, 50 chunks, SGD
  86. eec0981a 14:06:12 sota_crawler Add short TTT experiment: no-EMA, 50 chunks, SGD
  87. d97ad7db 14:23:24 parameter-golf-lab DG: use Flash Attention for designator matching
  88. b6374c09 14:27:02 parameter-golf-lab Add XSA-all(11) + short TTT run script
  89. d36c1595 14:27:02 parameter-golf-lab Add XSA-all(11) + short TTT run script
  90. 27833024 14:27:02 parameter-golf-lab Add XSA-all(11) + short TTT run script
  91. b6374c09 14:27:02 sota_crawler Add XSA-all(11) + short TTT run script
  92. d36c1595 14:27:02 sota_crawler Add XSA-all(11) + short TTT run script
  93. 27833024 14:27:02 sota_crawler Add XSA-all(11) + short TTT run script
  94. 03a4ef6f 14:37:25 parameter-golf-lab Record: v7 GPTQ + Short TTT — val_bpb 1.1207 (seed 1337)
  95. 4185795d 14:37:25 parameter-golf-lab Record: v7 GPTQ + Short TTT — val_bpb 1.1207 (seed 1337)
  96. eac2e723 14:37:25 parameter-golf-lab Record: v7 GPTQ + Short TTT — val_bpb 1.1207 (seed 1337)
  97. 03a4ef6f 14:37:25 sota_crawler Record: v7 GPTQ + Short TTT — val_bpb 1.1207 (seed 1337)
  98. 4185795d 14:37:25 sota_crawler Record: v7 GPTQ + Short TTT — val_bpb 1.1207 (seed 1337)
  99. eac2e723 14:37:25 sota_crawler Record: v7 GPTQ + Short TTT — val_bpb 1.1207 (seed 1337)
  100. 5a5bbad3 14:39:07 parameter-golf-lab recursive weight-sharing transformer compression shim on SwiGLU SOTA (1.0763 model)
  101. 874237b6 14:39:07 parameter-golf-lab recursive weight-sharing transformer compression shim on SwiGLU SOTA (1.0763 model)
  102. 9fd191d5 14:39:07 parameter-golf-lab recursive weight-sharing transformer compression shim on SwiGLU SOTA (1.0763 model)
  103. 5a5bbad3 14:39:07 sota_crawler recursive weight-sharing transformer compression shim on SwiGLU SOTA (1.0763 model)
  104. 874237b6 14:39:07 sota_crawler recursive weight-sharing transformer compression shim on SwiGLU SOTA (1.0763 model)
  105. 9fd191d5 14:39:07 sota_crawler recursive weight-sharing transformer compression shim on SwiGLU SOTA (1.0763 model)
  106. 57314a2a 14:50:31 parameter-golf-lab Add v7 smooth run: proper warmdown + XSA-all
  107. 47b7700f 14:50:31 parameter-golf-lab Add v7 smooth run: proper warmdown + XSA-all
  108. 57314a2a 14:50:31 sota_crawler Add v7 smooth run: proper warmdown + XSA-all
  109. 47b7700f 14:50:31 sota_crawler Add v7 smooth run: proper warmdown + XSA-all
  110. a79acfdf 14:51:37 parameter-golf-lab Reduce bigram buckets 2048→1792 for XSA-11 size fit
  111. da0f71aa 14:51:37 parameter-golf-lab Reduce bigram buckets 2048→1792 for XSA-11 size fit
  112. a79acfdf 14:51:37 sota_crawler Reduce bigram buckets 2048→1792 for XSA-11 size fit
  113. da0f71aa 14:51:37 sota_crawler Reduce bigram buckets 2048→1792 for XSA-11 size fit
  114. 510b1bc6 14:59:50 parameter-golf-lab DG v3: hardcoded depth schedule replaces learned β
  115. 6046116c 15:09:02 parameter-golf-lab recursive weight-sharing transformer stacked: LeakyReLU(0.5)² + VRL + decoder_lr_mult=2.0
  116. 7efd4e32 15:09:02 parameter-golf-lab recursive weight-sharing transformer stacked: LeakyReLU(0.5)² + VRL + decoder_lr_mult=2.0
  117. 6046116c 15:09:02 sota_crawler recursive weight-sharing transformer stacked: LeakyReLU(0.5)² + VRL + decoder_lr_mult=2.0
  118. 7efd4e32 15:09:02 sota_crawler recursive weight-sharing transformer stacked: LeakyReLU(0.5)² + VRL + decoder_lr_mult=2.0
  119. f96d349e 15:16:28 parameter-golf-lab Add research batch script + GPTQ env var tuning
  120. f523db4a 15:16:28 parameter-golf-lab Add research batch script + GPTQ env var tuning
  121. f96d349e 15:16:28 sota_crawler Add research batch script + GPTQ env var tuning
  122. f523db4a 15:16:28 sota_crawler Add research batch script + GPTQ env var tuning
  123. 8ff3e0ef 15:19:40 parameter-golf-lab Record: LeakyReLU² + Legal TTT + Parallel Muon — val_bpb 1.1194 (3-seed mean)
  124. 8ff3e0ef 15:19:40 sota_crawler Record: LeakyReLU² + Legal TTT + Parallel Muon — val_bpb 1.1194 (3-seed mean)
  125. 8ff3e0ef 15:19:40 SOTA_FINAL Record: LeakyReLU² + Legal TTT + Parallel Muon — val_bpb 1.1194 (3-seed mean)
  126. 9fec4314 15:21:45 parameter-golf-lab Add DG Attention non-record submission (val_bpb=1.1898)
  127. b3c24d97 15:27:47 parameter-golf-lab Clean PR: remove dev artifacts and track_10min_16mb files
  128. 9dd29bea 15:29:03 parameter-golf-lab Sanitize: remove all illegal TTT from recursive weight-sharing transformer scripts
  129. ddb28362 15:29:03 parameter-golf-lab Sanitize: remove all illegal TTT from recursive weight-sharing transformer scripts
  130. 9dd29bea 15:29:03 sota_crawler Sanitize: remove all illegal TTT from recursive weight-sharing transformer scripts
  131. ddb28362 15:29:03 sota_crawler Sanitize: remove all illegal TTT from recursive weight-sharing transformer scripts
  132. 736d1bd9 15:41:01 parameter-golf-lab Add recursive weight-sharing transformer SwiGLU sweep: 9 configs across 3 batches
  133. 6f8be1f3 15:41:01 parameter-golf-lab Add recursive weight-sharing transformer SwiGLU sweep: 9 configs across 3 batches
  134. 736d1bd9 15:41:01 sota_crawler Add recursive weight-sharing transformer SwiGLU sweep: 9 configs across 3 batches
  135. 6f8be1f3 15:41:01 sota_crawler Add recursive weight-sharing transformer SwiGLU sweep: 9 configs across 3 batches
  136. 214fa9e8 15:51:52 parameter-golf-lab EMERGENCY: Disable all illegal TTT across entire project
  137. 496705c9 15:51:52 parameter-golf-lab EMERGENCY: Disable all illegal TTT across entire project
  138. 214fa9e8 15:51:52 sota_crawler EMERGENCY: Disable all illegal TTT across entire project
  139. 496705c9 15:51:52 sota_crawler EMERGENCY: Disable all illegal TTT across entire project
  140. 1a01dc77 15:56:27 parameter-golf-lab CRITICAL: Remove illegal TTT from train_gpt_swiglu.py
  141. 560db561 15:56:27 parameter-golf-lab CRITICAL: Remove illegal TTT from train_gpt_swiglu.py
  142. 1a01dc77 15:56:27 sota_crawler CRITICAL: Remove illegal TTT from train_gpt_swiglu.py
  143. 560db561 15:56:27 sota_crawler CRITICAL: Remove illegal TTT from train_gpt_swiglu.py
  144. 605f965a 16:03:38 parameter-golf-lab DG v4: prev-token baseline, single projection, depth-skip
  145. bc4f3420 16:24:27 parameter-golf-lab Add SwiGLU recursive weight-sharing transformer compression sweep (6 configs)
  146. ebf06e18 16:24:27 parameter-golf-lab Add SwiGLU recursive weight-sharing transformer compression sweep (6 configs)
  147. bc4f3420 16:24:27 sota_crawler Add SwiGLU recursive weight-sharing transformer compression sweep (6 configs)
  148. ebf06e18 16:24:27 sota_crawler Add SwiGLU recursive weight-sharing transformer compression sweep (6 configs)
  149. c2da2f2f 16:32:53 parameter-golf-lab SwiGLU F1: VRL + LeakyReLU(0.5)² + legal TTT + int8 attn.proj + seq2048
  150. 4a629336 16:32:53 parameter-golf-lab SwiGLU F1: VRL + LeakyReLU(0.5)² + legal TTT + int8 attn.proj + seq2048
  151. c2da2f2f 16:32:53 sota_crawler SwiGLU F1: VRL + LeakyReLU(0.5)² + legal TTT + int8 attn.proj + seq2048
  152. 4a629336 16:32:53 sota_crawler SwiGLU F1: VRL + LeakyReLU(0.5)² + legal TTT + int8 attn.proj + seq2048
  153. 4e5de0b7 16:39:38 parameter-golf-lab Fix DDP: find_unused_parameters=True for VRL block 0
  154. 77a8a188 16:39:38 parameter-golf-lab Fix DDP: find_unused_parameters=True for VRL block 0
  155. 4e5de0b7 16:39:38 sota_crawler Fix DDP: find_unused_parameters=True for VRL block 0
  156. 77a8a188 16:39:38 sota_crawler Fix DDP: find_unused_parameters=True for VRL block 0
  157. 2aae6b0a 16:45:42 parameter-golf-lab Add GPTQ re-quantization sweep (no training, ~2 min)
  158. d66040ee 16:45:42 parameter-golf-lab Add GPTQ re-quantization sweep (no training, ~2 min)
  159. 2aae6b0a 16:45:42 sota_crawler Add GPTQ re-quantization sweep (no training, ~2 min)
  160. d66040ee 16:45:42 sota_crawler Add GPTQ re-quantization sweep (no training, ~2 min)
  161. ffc909b0 16:51:09 parameter-golf-lab Save Gold Standard (GS): v7 GPTQ 1.1206 BPB — 3 copies
  162. 3e624068 16:51:09 parameter-golf-lab Save Gold Standard (GS): v7 GPTQ 1.1206 BPB — 3 copies
  163. ffc909b0 16:51:09 sota_crawler Save Gold Standard (GS): v7 GPTQ 1.1206 BPB — 3 copies
  164. 3e624068 16:51:09 sota_crawler Save Gold Standard (GS): v7 GPTQ 1.1206 BPB — 3 copies
  165. 9fefc754 16:52:47 parameter-golf-lab Restore original train_gpt_swiglu.py, new F1 variant as separate file
  166. 4b9ed8b2 16:52:47 parameter-golf-lab Restore original train_gpt_swiglu.py, new F1 variant as separate file
  167. 9fefc754 16:52:47 sota_crawler Restore original train_gpt_swiglu.py, new F1 variant as separate file
  168. 4b9ed8b2 16:52:47 sota_crawler Restore original train_gpt_swiglu.py, new F1 variant as separate file
  169. 0340c66f 17:00:01 parameter-golf-lab Log SwiGLU F1 run: 1.1208 BPB (20.6MB, needs compression)
  170. 16db94f0 17:00:01 parameter-golf-lab Log SwiGLU F1 run: 1.1208 BPB (20.6MB, needs compression)
  171. 0340c66f 17:00:01 sota_crawler Log SwiGLU F1 run: 1.1208 BPB (20.6MB, needs compression)
  172. 16db94f0 17:00:01 sota_crawler Log SwiGLU F1 run: 1.1208 BPB (20.6MB, needs compression)
  173. 139d573e 17:13:03 parameter-golf-lab Fix pre-TTT BPB, TTT gains, and steps to match logs exactly
  174. 139d573e 17:13:03 sota_crawler Fix pre-TTT BPB, TTT gains, and steps to match logs exactly
  175. 139d573e 17:13:03 SOTA_FINAL Fix pre-TTT BPB, TTT gains, and steps to match logs exactly
  176. b08d72a9 17:22:25 parameter-golf-lab Fix author attributions: PR #493 @parinzee, PR #461 @Christopher-Lee-McClendon
  177. b08d72a9 17:22:25 sota_crawler Fix author attributions: PR #493 @parinzee, PR #461 @Christopher-Lee-McClendon
  178. b08d72a9 17:22:25 SOTA_FINAL Fix author attributions: PR #493 @parinzee, PR #461 @Christopher-Lee-McClendon
  179. b82c50dd 17:36:14 parameter-golf-lab Update README leaderboard with merged records
  180. b82c50dd 17:36:14 sota_crawler Update README leaderboard with merged records
  181. b82c50dd 17:36:14 SOTA_FINAL Update README leaderboard with merged records
  182. d74c0b58 17:38:56 parameter-golf-lab Use GitHub usernames in new leaderboard rows
  183. d74c0b58 17:38:56 sota_crawler Use GitHub usernames in new leaderboard rows
  184. d74c0b58 17:38:56 SOTA_FINAL Use GitHub usernames in new leaderboard rows
  185. 8a77849b 17:41:35 parameter-golf-lab Describe leaderboard entries by base-run diff
  186. 8a77849b 17:41:35 sota_crawler Describe leaderboard entries by base-run diff
  187. 8a77849b 17:41:35 SOTA_FINAL Describe leaderboard entries by base-run diff
  188. c7311ad2 18:00:18 parameter-golf-lab DG v4.1: eliminate second matmul via linearity trick
  189. 89914646 20:33:11 parameter-golf-lab Add TTT calibration sweep: 11-config grid to close 1.11 BPB gap
  190. 0efae917 20:33:11 parameter-golf-lab Add TTT calibration sweep: 11-config grid to close 1.11 BPB gap
  191. 89914646 20:33:11 sota_crawler Add TTT calibration sweep: 11-config grid to close 1.11 BPB gap
  192. 0efae917 20:33:11 sota_crawler Add TTT calibration sweep: 11-config grid to close 1.11 BPB gap
  193. a89a1426 21:23:53 parameter-golf-lab Add GPTQ re-quant sweep for GS v7 — no retraining needed
  194. d6dad03c 21:23:53 parameter-golf-lab Add GPTQ re-quant sweep for GS v7 — no retraining needed
  195. a89a1426 21:23:53 sota_crawler Add GPTQ re-quant sweep for GS v7 — no retraining needed
  196. d6dad03c 21:23:53 sota_crawler Add GPTQ re-quant sweep for GS v7 — no retraining needed
  197. 513fd618 21:43:51 parameter-golf-lab Update recursive weight-sharing transformer PR with full experiment history and research directions
  198. 6abf4436 21:43:51 parameter-golf-lab Update recursive weight-sharing transformer PR with full experiment history and research directions
  199. 513fd618 21:43:51 sota_crawler Update recursive weight-sharing transformer PR with full experiment history and research directions
  200. 6abf4436 21:43:51 sota_crawler Update recursive weight-sharing transformer PR with full experiment history and research directions
  201. d0c4271d 22:00:21 parameter-golf-lab Rewrite recursive weight-sharing transformer PR: honest framing, clean architecture/training separation
  202. d0c4271d 22:00:21 sota_crawler Rewrite recursive weight-sharing transformer PR: honest framing, clean architecture/training separation
  203. da89c9ee 22:13:07 parameter-golf-lab Submission: XSA-11 + GPTQ block64/percdamp002
  204. 8fbd41db 22:13:07 parameter-golf-lab Submission: XSA-11 + GPTQ block64/percdamp002
  205. d6748136 22:13:07 parameter-golf-lab Submission: XSA-11 + GPTQ block64/percdamp002
  206. da89c9ee 22:13:07 sota_crawler Submission: XSA-11 + GPTQ block64/percdamp002
  207. 8fbd41db 22:13:07 sota_crawler Submission: XSA-11 + GPTQ block64/percdamp002
  208. 0100f8e9 23:15:01 parameter-golf-lab XSA-11 + GPTQ block64/percdamp002 — 3-seed mean 1.1208 BPB
  209. 0100f8e9 23:15:01 sota_crawler XSA-11 + GPTQ block64/percdamp002 — 3-seed mean 1.1208 BPB
  210. 303192e9 23:16:37 parameter-golf-lab XSA-11 + GPTQ block64/percdamp002 — 3-seed mean val_bpb 1.1208
  211. 303192e9 23:16:37 sota_crawler XSA-11 + GPTQ block64/percdamp002 — 3-seed mean val_bpb 1.1208
  212. 54f14bd9 00:37:50 parameter-golf-lab Add int5 sizing sweep — test how many params fit at int5 vs int6
  213. 44ae7a85 00:40:58 parameter-golf-lab Add int5 sizing sweep — test how many params fit at int5
  214. 44ae7a85 00:40:58 sota_crawler Add int5 sizing sweep — test how many params fit at int5
  215. 7bbb09cd 00:50:04 parameter-golf-lab Int5 test: 33.6M params (MHA 8/8, MLP 3.5x) with int5 GPTQ
  216. 7bbb09cd 00:50:04 sota_crawler Int5 test: 33.6M params (MHA 8/8, MLP 3.5x) with int5 GPTQ
  217. bd415d5c 01:14:49 parameter-golf-lab Add 576+ single-file runner and aggressive edge test launchers
  218. bd415d5c 01:14:49 sota_crawler Add 576+ single-file runner and aggressive edge test launchers
  219. 30a61762 02:47:29 parameter-golf-lab Add TTT sweep infra + Vast.ai pipeline + int5mix run record
  220. 30a61762 02:47:29 sota_crawler Add TTT sweep infra + Vast.ai pipeline + int5mix run record
  221. f1b68023 02:48:19 parameter-golf-lab Add RunPod setup script for TTT calibration sweep
  222. f1b68023 02:48:19 sota_crawler Add RunPod setup script for TTT calibration sweep
  223. 8f94780c 03:10:53 parameter-golf-lab Add Micro Crawler architecture — asymmetric fractal cadence test cadence test with flat+crawler split
  224. 8f94780c 03:10:53 sota_crawler Add Micro Crawler architecture — asymmetric fractal cadence test cadence test with flat+crawler split
  225. b3b408bb 03:13:15 parameter-golf-lab Update micro crawler to balanced 4f+2cx2 config at dim=640
  226. b3b408bb 03:13:15 sota_crawler Update micro crawler to balanced 4f+2cx2 config at dim=640
  227. 65f42e86 03:41:23 parameter-golf-lab Add crawler cadence + per-loop GPTQ to H100 micro crawler
  228. 65f42e86 03:41:23 sota_crawler Add crawler cadence + per-loop GPTQ to H100 micro crawler
  229. c9e2aff8 03:53:47 parameter-golf-lab Cadence N/N/N/N/C + activation-aware per-loop GPTQ blending
  230. c9e2aff8 03:53:47 sota_crawler Cadence N/N/N/N/C + activation-aware per-loop GPTQ blending
  231. ee034506 03:59:51 parameter-golf-lab Pre-flight: backup scripts + pod setup for micro crawler 8xH100
  232. ee034506 03:59:51 sota_crawler Pre-flight: backup scripts + pod setup for micro crawler 8xH100
  233. 5a1e8b76 04:10:39 parameter-golf-lab Recursive cadence: ramp N count as LR warms down
  234. 5a1e8b76 04:10:39 sota_crawler Recursive cadence: ramp N count as LR warms down
  235. 81ed4752 04:21:36 parameter-golf-lab recursive weight-sharing transformer v2: trigram hash + lr_mul compile-spike fix
  236. 81ed4752 04:21:36 sota_crawler recursive weight-sharing transformer v2: trigram hash + lr_mul compile-spike fix
  237. de856119 04:24:00 parameter-golf-lab Fix lr_mul compile spike + shrink dim 640→624 for artifact budget
  238. de856119 04:24:00 sota_crawler Fix lr_mul compile spike + shrink dim 640→624 for artifact budget
  239. 84f32ff5 04:25:12 parameter-golf-lab Freeze run1 scripts — 1.1377 sliding window (broken LR, dim=640)
  240. 84f32ff5 04:25:12 sota_crawler Freeze run1 scripts — 1.1377 sliding window (broken LR, dim=640)
  241. a35820ee 04:25:47 parameter-golf-lab Fix dim 624→630 (must be divisible by num_heads=10)
  242. a35820ee 04:25:47 sota_crawler Fix dim 624→630 (must be divisible by num_heads=10)
  243. 3cfebf8d 04:26:51 parameter-golf-lab Fix dim 630→620 (head_dim must be even for RoPE)
  244. 3cfebf8d 04:26:51 sota_crawler Fix dim 630→620 (head_dim must be even for RoPE)
  245. 357b9aff 04:29:14 parameter-golf-lab Fix dim 620→560 (FA3 needs head_dim multiple of 8: 560/10=56)
  246. 357b9aff 04:29:14 sota_crawler Fix dim 620→560 (FA3 needs head_dim multiple of 8: 560/10=56)
  247. 61dce75b 04:31:22 parameter-golf-lab Add quantfix training script: learned scales, earlier QAT, selective int8
  248. 61dce75b 04:31:22 sota_crawler Add quantfix training script: learned scales, earlier QAT, selective int8
  249. 94cd3eba 04:39:29 parameter-golf-lab Run2: GPTQ Hessian quant + dim=640 + trigram_vocab=2048
  250. 94cd3eba 04:39:29 sota_crawler Run2: GPTQ Hessian quant + dim=640 + trigram_vocab=2048
  251. 9d50ab72 04:42:29 parameter-golf-lab Launcher: dim=640, trigram_vocab=2048 (fit budget, test lr_mul fix only)
  252. 9d50ab72 04:42:29 sota_crawler Launcher: dim=640, trigram_vocab=2048 (fit budget, test lr_mul fix only)
  253. 1a6ab1b3 04:43:10 parameter-golf-lab Keep trigram_vocab=8192 to match run1 — isolate lr_mul fix only
  254. 1a6ab1b3 04:43:10 sota_crawler Keep trigram_vocab=8192 to match run1 — isolate lr_mul fix only
  255. bcccc19d 04:53:13 parameter-golf-lab Run3: Self-referential crawler with deliberation gate
  256. bcccc19d 04:53:13 sota_crawler Run3: Self-referential crawler with deliberation gate
DATE2026-03-23
DAY006
PRIMARY METRICVAL_BPB
TESTS RUN22 unique (22 valid / 0 TTT leakage)
ARCHITECTURE FAMILYLEADERBOARD TRANSFORMER VARIANTS (×11) · CROSS-SEQUENCE ATTENTION VARIANT (×3) · STATE-SPACE TTT VARIANT (×2) · POST-TRAINING QUANTIZATION CALIBRATION (×2) · RECURSIVE WEIGHT-SHARING TRANSFORMER (×1) · SLIDING-WINDOW EVALUATION VARIANT (×1) · +2 MORE (×2)
CLASSIFICATIONEXPERIMENTAL 82% (18)/NEURAL 18% (4)

SECTION 1.0 — TRAJECTORY

VAL_BPB for 2026-03-23 panel 1 of 2
VAL_BPB for 2026-03-23 panel 2 of 2
VAL_BPB SPREADnminmedianmaxdownload
LEADERBOARD TRANSFORMER VARIANTS111.06721.12081.4138
CROSS-SEQUENCE ATTENTION VARIANT31.12081.12082.0644
STATE-SPACE TTT VARIANT21.12011.12621.1322
POST-TRAINING QUANTIZATION CALIBRATION21.12061.12101.1215download
RECURSIVE WEIGHT-SHARING TRANSFORMER11.14781.14781.1478download
SLIDING-WINDOW EVALUATION VARIANT11.13771.13771.1377
COMPACT TRANSFORMER RECIPE11.12071.12071.1207
CRAWLER ABLATION CONTROL SET11.37661.37661.3766
ALL VALID221.06721.12092.0644

SECTION 2.0 — TEST RUNS

EXPERIMENTAL TEST RUNS — 18

#TIMEHWABLATIONVAL_BPBSTATUSΔ %OPEN
204:21recursive weight-sharing transformer v2: trigram hash + lr_mul compile-spike …1.1478OK↑2.4%
details
304:25Freeze run1 scripts — 1.1377 sliding window (broken LR, dim=640) (×2)1.1377OK↑1.5%
details
405:54Add AdamW TTT option — PR #462 shows 5x better TTT gain vs SGD (×7)1.0672OK↓4.8%
details
606:26SwiGLU + GPTQ: PR #462 architecture with our quant improvements (×6)1.0672OK↓4.8%
details
707:12Log session results: 1.0763 SwiGLU (size problem), 1.1215 v7 (submitted) (×6)1.0763OK↓4.0%
details
814:358×H100run_s1337.log (×3)1.1201OK↓0.1%
training_log 1
training_log 2
training_log 3
1016:241×GB10A_bigram1536_xsa11.log2.0644OK↑84.2%
training_log
1116:51Save Gold Standard (GS): v7 GPTQ 1.1206 BPB — 3 copies (×4)1.1206OK↓0.0%
details
1317:031×GB10S1_loops3_baseline.log1.3766OK↑22.8%
training_log
1417:391×GB10S2_loops4.log1.4058OK↑25.4%
training_log
1518:171×GB10S3_loops5.log1.4138OK↑26.1%
training_log
1618:531×GB10S4_share3_loops4.log1.4076OK↑25.6%
training_log
1719:061×GB10S5_loops4_bigram4096.log1.4123OK↑26.0%
training_log
1822:178×H100 (runpod)submit_s1337_20260323_221742.log1.1203OK↓0.1%
training_log
1922:388×H100 (runpod)submit_s42_20260323_223824.log1.1211OK↑0.0%
training_log
2022:568×H100 (runpod)submit_s7_20260323_225635.log1.1205OK↓0.0%
training_log
2123:15XSA-11 + GPTQ block64/percdamp002 — 3-seed mean 1.1208 BPB (×2)1.1208OK↓0.0%
details
2223:16XSA-11 + GPTQ block64/percdamp002 — 3-seed mean val_bpb 1.1208 (×2)1.1208OK↓0.0%
details

NEURAL TEST RUNS — 4

#TIMEHWABLATIONVAL_BPBSTATUSΔ %OPEN
102:47Add TTT sweep infra + Vast.ai pipeline + int5mix run record (×2)1.1322OK↑1.0%
details
506:02Record: GPTQ + Early QAT + TTT EMA — 3-seed mean val_bpb 1.1215 (×3)1.1215OK↑0.0%
details
914:37Record: v7 GPTQ + Short TTT — val_bpb 1.1207 (seed 1337) (×6)1.1207OK↓0.0%
details
1217:00Log SwiGLU F1 run: 1.1208 BPB (20.6MB, needs compression) (×4)1.1208OK↓0.0%
details