Phase 3: re-draft between judge passes
Question. Today the slots the judge rejects are re-masked and the 30B fills them itself, a few per pass, often one forced token per pass. Does handing those slots back to the drafter cut judge passes without losing accuracy?
Method, per 32-token block
- SDAR-1.7B drafts every slot (8 steps), exactly as today.
- Judge pass (30B + the 128M LoRA): commit where p > 0.95 (C); keep the draft where the judge agrees but is unsure (t); re-mask where it disagrees (M). If nothing clears 0.95, the most confident slot is committed (forced), as today.
- After every K judge passes, the drafter re-fills only the M slots. It sees only verified tokens: the prompt plus
the committed slots. The t slots are hidden as [MASK] in its input and keep their token in the block.
Steps = ceil(#M / 4), at most 8 (the first draft's rate). Logged as drafter kind
redraft. - Repeat 2–3 until every slot is committed. Only the judge commits, so every token's last step is a judge pass.
Setups (both on all 7 benchmarks)
| Setup | Rule | Flag |
|---|---|---|
| rd1 | re-draft after every judge pass | --full-redraft-every 1 |
| rd2 | re-draft after every 2 judge passes | --full-redraft-every 2 |
Benchmarks: HumanEval 164, MBPP 257, GSM8K-500, IFEval-541, MATH-500, GPQA-Diamond 198, AIME 90.
Identical to every earlier run: run_da_decode.py (PyTorch, no SGLang); JetLM/SDAR-30B-A3B-Chat-b32 @ c351bbc and
SDAR-1.7B-Chat-b32 @ 2c83c05; block 32; threshold 0.95; tau-commit 0.95; s8 first draft; --judge full; Phase 1's 128M
judge (Arushhh/c1b-judge-lora-sdar30b-b32, main_128m/tok_000128000000, weights hash 35805974); no expert pool (no
DES); greedy; the same benchmark files and length limits (research_protocol.GEN_LEN); 8 shards, one per GPU;
--score --no-trace; the Phase 2 environment (/venv/p2: torch 2.11.0 cu128, transformers 4.54.1); 8× RTX PRO 6000.
Run
bash code/phase3/run_p3.sh 2>&1 | tee -a logs/phase3/run.log # inside tmux; resumes if re-run
It runs setup (setup_p3.sh: Phase 2's environment and downloads, the 128M judge, Phase 1's reference records), the
checks, rd1 and rd2 on each benchmark (alternating which goes first), the analysis and a backup after every benchmark.
Checks (kept minimal): CPU tests; stock on 16 GSM8K-500 items must match the recorded outputs token for token (16/16); SDAR-1.7B + judge on 16 HumanEval items vs Phase 1's records (at least 12/16 identical).
Outputs
logs/phase3/decode/<rd1|rd2>/<benchmark>/: per-item records;jobs/: shard logs;setup/: setup, checksresults/phase3/TABLE.mdandmetrics.json(also committed to this branch)- Hugging Face dataset
Arushhh/phase3-redraft(public): logs, results and the code that ran, synced after every benchmark
Reading the results
- Compare with today's judge and stock (Phase 1 / round-3 records): accuracy, passes per block, expert loads per token, drafter forwards per token. These don't depend on the machine.
- Tokens per second compares rd1 with rd2 only (same box). Earlier timings come from other machines.
- Speed needs fewer judge passes per block than today (5.9 on HumanEval up to 15.6 on IFEval). One drafter step costs about 9.5 ms and one judge pass about 250 ms (Phase 1 box), so each re-draft pays for itself if it saves about 0.04 judge passes per drafter step.
- Diagnostics per setup: re-draft rounds and slots per block; how often a re-drafted token equals the judge's top choice from the pass before; how often it repeats the rejected draft; next-slot copy rate at draft and at mask slots.
Time (US box, $11.75/hr)
Setup about 1 h; each setup about 2.8 h on the 7 benchmarks; about 6.5 h in all, up to 8.5 h if the box's CPU slows the decode loop (about $80–100).
Code changes (all off by default: every earlier setting runs exactly as before)
code/draftalign/da_decode.py:judge_full_block(redraft_every=..., diag=...),copy_diag, new record fields (written only when the options are on)code/draftalign/drafter.py:draft_block(static_steps=..., kind=...)code/draftalign/run_da_decode.py:--full-redraft-every,--full-redraft-rate,--judge-diag; cell suffix_rd<K>code/tests/test_da_decode_redraft.py