Commit Graph

  • e6be33aa53 doc: add SFT length-filter rationale — high PPL + high variance from A.3 IFD figure master ViperEkura 2026-07-24 06:37:35 +08:00
  • 0a1d0573ae doc: add SFT length filter (15-token floor) with IFD bias rationale ViperEkura 2026-07-24 06:27:10 +08:00
  • 0c2bc916f2 fix: loss_compare caption token count (20B -> 5B) ViperEkura 2026-07-24 06:21:52 +08:00
  • 4e70e827ff remove: drop ckpt_weight_density_per_run.png figure ViperEkura 2026-07-24 06:20:08 +08:00
  • a7bbc7b29f fix: align SFT/DPO figures and text with actual training data ViperEkura 2026-07-24 06:15:42 +08:00
  • 4b37f289c0 fix: attribute muon-25bt larger std to Muon characteristic, not training duration alone ViperEkura 2026-07-20 11:59:09 +08:00
  • a83555f326 fix: align conclusion, abstract, introduction with body data and full paper scope ViperEkura 2026-07-20 11:51:48 +08:00
  • d93ff48320 docs: fix storage backends (2->3) and per-field length filter ViperEkura 2026-07-20 11:27:39 +08:00
  • da0f536526 Update DPO iteration count: 1,500 -> 3,000 ViperEkura 2026-07-20 03:59:55 +08:00
  • c775a2b3e0 Update DPO metrics figure ViperEkura 2026-07-20 03:59:04 +08:00
  • 3f0ff911a8 Fix lingering cosine->WSD in Conclusion; update AGENTS.md with checkpoint labels and scheduler rules ViperEkura 2026-07-19 15:30:49 +08:00
  • e149997200 Add DPO data generation & training; fix checkpoint labels; switch to WSD scheduler; update optimizer ablations in abstract/conclusion ViperEkura 2026-07-19 15:25:42 +08:00
  • 6c2e04a86f Add DPO training subsection with dpo_metrics figure and Rafailov et al. citation ViperEkura 2026-07-19 15:01:23 +08:00
  • 411354eeb1 Update sft_metrics.png ViperEkura 2026-07-18 00:52:52 +08:00
  • a0c39601a0 Add effective rank analysis, weight distribution evolution, Muon/SFT training figures ViperEkura 2026-07-17 23:05:21 +08:00
  • bfc8ff6098 Fix training budget (19k steps / ~20B tokens), add Muon optimizer, correct variance scaling math, move figures to §Training Config, tighten abstract ViperEkura 2026-07-08 22:49:43 +08:00
  • 4a143b056d Tighten abstract, add Muon optimizer mention, update ckpt_comparison figure ViperEkura 2026-07-08 22:27:33 +08:00
  • 5726bf6c84 Add optimizer/initialization comparison figure (ckpt_comparison) ViperEkura 2026-07-08 21:59:56 +08:00
  • ff13a45d95 Restructure data pipeline: replace IFD analysis with length bias appendix ViperEkura 2026-07-05 18:59:29 +08:00
  • 2d52a00149 Update ifd_loss_ratio_density.png ViperEkura 2026-07-05 12:38:43 +08:00
  • f2c2b159ba Add IFD loss ratio density analysis, weight distribution appendix, and rewrite abstract ViperEkura 2026-07-05 12:31:40 +08:00
  • 459523f0e4 Fix IFD citation: correct author list and venue (NAACL 2024, not NeurIPS) ViperEkura 2026-07-05 00:24:43 +08:00
  • 4a7bf7ac3f Add GQA/SwiGLU/RoPE formula blocks, Alpaca-GPT4 citation, and improve IFD-Loss Ratio interpretation ViperEkura 2026-07-05 00:17:09 +08:00
  • 5c915514b3 Add correlation analysis interpretation to IFD bias section ViperEkura 2026-07-04 23:45:55 +08:00
  • bc8d4066c1 Add IFD analysis: density/length-bias figures and appendix ViperEkura 2026-07-04 23:33:12 +08:00
  • 9534316bac Fix duplicate paragraph and misleading table caption ViperEkura 2026-07-04 22:39:16 +08:00
  • b40fcc86c7 Refine abstract, switch to Times font, remove date, fix overfull hboxes with microtype ViperEkura 2026-07-04 22:31:45 +08:00
  • dd412b3c9c Add IFD analysis, appendix, and improve abstract ViperEkura 2026-07-04 22:25:18 +08:00
  • 73cb02f0aa fix training hyperparams to match pretraining script: lr 1.5e-4, wd 0.1, warmup 0.02 ViperEkura 2026-06-28 16:57:30 +08:00
  • adc284920a add Alembic SFT data cleaning section with MinHash algorithm and fix bibliography order ViperEkura 2026-06-28 16:52:11 +08:00
  • 13089b002a fix training config: grad_accum 32, remove unused validation & activation ckpt ViperEkura 2026-06-28 15:35:15 +08:00
  • d3fbe0ecc6 refactor: numerical stability analysis with residual scaling comparison ViperEkura 2026-06-28 14:36:12 +08:00
  • 3db1096c78 fix: add caption package, fix overfull hbox, correct RoFormer author names ViperEkura 2026-06-25 16:53:26 +08:00
  • 6f9278b35b Initial commit: BF16 weight lock-in paper for 1.2B Transformer training with AstrAI ViperEkura 2026-06-25 15:51:05 +08:00