vh
d710e56aca
Split the 53 over-threshold dated log entries (Recent decisions, Tried and abandoned) into per-entry persistent-memory.d/<slug>.md detail files, leaving one-line pointers in the index; the 10 short entries stay inline. Startup index drops 60,527 -> 25,256 bytes (492 -> 270 lines); entry bodies move verbatim to on-demand detail files, so a fresh session loads ~25 KB instead of ~60 KB and pulls a detail file only when its pointer is relevant. Top matter (Repo purpose, Tools & conventions, Current state) is unchanged; both archival back-references preserved. CLAUDE.md persistent-memory section now documents the index<->detail read discipline (read the index, pull details on demand, never bulk-read the dir, commit both together). Auto-archival still held every dated entry back (all <30 days old); the July burst begins aging past the 30-day guard ~2026-07-31.
2.1 KiB
2.1 KiB
[2026-07-08]DPO was silently running 3 epochs (harness gap) → KILLED at epoch 1.2, retargeted to 0.3 epochs (operator call). Root cause:DpoConfighad NOepochsfield +_dpo_config_kwargsdidn't passnum_train_epochs→ DPO fell through to trl DPOConfig's default 3.0 (SFT correctly pins 1 via SftConfig.epochs- _sft_config_kwargs). Objective SATURATED by
epoch 0.27 (loss0, grad0, acc 1.0, margins27 flat — the off-policy qwopus rejected pairs are trivially separable), so epochs ~0.3→3 were pure over-optimization + a ~6.5h outage. No mid-run checkpoint (save_steps 500 > total steps; save only at end) → killing lost the run. FIX (3 edits to deployed harness, mtf-dev to canonicalize):DpoConfig.epochs: float = 1(mirrors SftConfig, float for fractions);_dpo_config_kwargsnow passes"num_train_epochs": cfg.epochs; recipedpo.epochs: 0.3. GPU-free verified (dpo.epochs=0.3 → num_train_epochs=0.3). Relaunched at 0.3 epoch (~30min precompute + ~12min train = ~45min). DONE + SANITY-CHECKED (exit 0, ~70min wall: ~30min fixed precompute + 45 steps @ ~51s/step; train_loss 1.4e-5 @ epoch 0.301). Freshdata/spike/dpo_adapter/adapter_model.safetensors(123MB) + checkpoint-45 banked. 3-way greedy sanity (base vs SFT vs DPO, via peft load + disable_adapter/set_adapter on GPU0): ALL THREE DISTINCT (base≠sft≠dpo) → full SFT→DPO pipeline applies end-to-end at inference. DPO 0.3ep is COHERENT, fluent, NOT degenerate (early-stop avoided over-optimization) but the quality delta on a neutral literary prompt is SUBTLE (DPO shares SFT's structure — it continues from it — with minor stylistic drift, arguably slightly MORE genre-clichéd). Verdict: mechanics proven, quality gain modest as predicted for 0.3ep on off-policy pairs; the real unlock remains on-policy rejected regen + on-domain (explicit E-RP) eval + the LitBench/holdout run. gen+rp RESTORED healthy. Next: serve fork (SGLang-finish vs merge) on the DPO adapter — same rekey_lora_for_vllm.py (zero-z) applies unchanged (mtf-dev confirmed).
- _sft_config_kwargs). Objective SATURATED by