Phase 2 drafter swap (no expert pool): code, results, data, weights in one place
Frozen SDAR-30B-A3B-Chat-b32 + the Phase 1 LoRA judge (128M-token corpus, weights hash 35805974), no expert pool,
nothing trained. Only the drafter changes: SDAR-1.7B (reference, Phase 1) vs CoDA-1.7B (F2_nopool) vs
Qwen3-0.6B diffusion (S1_nopool), on HumanEval 164, MBPP 257, GSM8K-500, IFEval-541, MATH-500, GPQA-D 198, AIME 90.
Get everything with one command
Tested 2026-10-05 from a clean environment without a token: 2,217 files, ~21 GB, about 7 min.
pip install -U huggingface_hub # provides the `hf` command (huggingface-cli was removed in huggingface_hub 1.x)
hf download Arushhh/phase2-drafter-swap --repo-type dataset --local-dir phase2_bundle # all ~21 GB; no token needed (public)
Layout
| path | what |
|---|---|
assets/results/TABLES.md, metrics.json |
every comparison table: accuracy, Δ vs SDAR-1.7B with 95% CI, experts/token, MoE-only / full-pipeline / drafter-only speedups, passes/block, draft tokens kept |
phase2/analysis/nopool/ |
the analysis: REPORT.md (dip? why? scope), IMPROVEMENT_PLAN.md, analysis.json, figures, scripts |
phase2/decode/F2_nopool/<set>/, phase2/decode/S1_nopool/<set>/ |
per-item records (*.jsonl), running metrics (*.running.jsonl), and experts.tar / judge.tar / traces.tar (per-forward expert bitmasks + router top-8, every judge read's probabilities, drafter confidences) |
phase2/plots/ |
all plots (also on W&B arushsharma536-spe-iit/phase2-drafter-swap) |
phase2/jobs/, phase2/*.log, phase2/setup/ |
job logs, launcher logs, environment freezes, setup checks |
assets/judge_128M_tok_000128000000/ |
the judge LoRA weights used (from Arushhh/c1b-judge-lora-sdar30b-b32) |
assets/ref_S0_SDAR1.7B_judge128M_nopool/<set>/ |
reference A: SDAR-1.7B + judge, no pool (Phase 1 records) |
assets/ref_stock_30B/ |
reference B: stock 30B decode records |
assets/datasets/ |
IFEval-541 (frozen, + sha256) and the manifest of the other benchmark datasets (public HF datasets) |
code/*.tar.gz |
the code at the commit that produced this (git archive); also on GitHub NoviceCoderInfinity/diffusion-moe-general-expert branch phase2-drafter-swap |
Not part of the results: phase2/decode/F2/, phase2/jobs/decode_F2* (non-nopool) and phase2/analysis/logprob/ are an
earlier run with the DES48 expert pool that was stopped at the user's request; ignore them.
Replicate on a new 8-GPU box
Unpack the code tarball (or clone the branch), put HF_TOKEN in .env, thenbash code/phase2/replicate_nopool.sh (see docs/PHASE2_NOPOOL_REPLICATE.md).
Models and drafters are public at pinned revisions: JetLM/SDAR-30B-A3B-Chat-b32 @ c351bbc, JetLM/SDAR-1.7B-Chat-b32 @ 2c83c05,
Salesforce/CoDA-v0-Instruct @ f2d40ce, dllm-hub/Qwen3-0.6B-diffusion-bd3lm-v0.1 @ 60f77bb.