jiayicheng/transfer-unit-catalyst

transfer-unit — Catalyst Data for Agent Distillation

Research codebase + full experimental record for the question: which auxiliary data makes a target task train FASTER (time-to-τ), and why?

Base model for everything: Qwen/Qwen3.5-4B (not included here — public weights).

What is in this repo

Path	Size	Contents
transfer-unit/tu/	800K	All library code (see below)
transfer-unit/scripts/	44K	Launch pipelines (frozen_pipe_v2.sh is the current one)
transfer-unit/configs/	144K	Every experiment config actually run
transfer-unit/tests/	16K	45 CPU acceptance tests (python3 tests/test_core.py)
transfer-unit/reports/	2.1M	All findings, figures, literature review, preregistrations
transfer-unit/data/*.jsonl.gz	7M	Donor rollout datasets, gzipped (11 tasks, token-ids + loss masks) + difficulty screenings
runs/<name>/	5M	Per-run results: curve.jsonl (pass@1 vs B-tokens), eval.jsonl, train_log.jsonl, sft_log.jsonl, rollouts.jsonl.gz (RL trace features, gzipped), config.json
runs/*.out	1.8M	Console logs (includes crash diagnostics worth keeping)

Deliberately excluded (regenerable / public): model checkpoints (~1.2 TB of .pt and ckpt_vllm/), the base model weights, the conda env, and runs/_invalid_prefix_bug/ (40 GB of runs invalidated by the sampling-seed bug documented in the reports).

Read the results here (in order)

1. reports/SPEED_THEORY.md — the main line: convergence-speed measurements (time-to-τ), the falsified theory v1, and the current rank-supplementation account.
2. reports/PREREGISTERED_PREDICTION.md, reports/PREREG_ROUND2.md — predictions locked BEFORE the corresponding training runs (round 1: 2 hits / 1 miss).
3. reports/LIT_REVIEW_speed.md — where this sits in the literature (gradient low-rank dynamics / multi-task conflict / fine-tuning prediction).
4. reports/LIT_REVIEW_catalyst.md — literature for the earlier data-attribute framing.
5. reports/FINAL_REPORT.md, reports/PILOT_RESULTS.md — the earlier endpoint-quality phase (superseded as a headline, kept for the controls: budget-matching, polarity).
6. reports/EXPERIMENT_BRIEF.md, reports/T0_novelty.md, reports/story_skeletons.md — original plan, novelty check, outcome-contingent paper skeletons.

Code map

tu/envs.py              multi-turn tool-use episode framework (action-block protocol)
tu/tasks/               11 synthetic, programmatically-verified tasks:
                        sql_query, pyfix, codegen, math_cot, math_hard, neutral_format,
                        logic_grid, story_qa, table_reason, translate, spec_write
tu/training/rollout.py  batched multi-turn rollout on vLLM with exact token accounting
tu/training/grpo.py     minimal GRPO: advantage-frozen trace filtering, span loss-masks
tu/offline/donate.py    collect donor rollouts (token ids + loss masks + trace features)
tu/offline/sft.py       SFT mixing experiments (B data + catalyst data, periodic ckpts)
tu/offline/eval_ckpts.py hot-swap checkpoint evaluator (~2.5 min/ckpt, engine loaded once)
tu/offline/export_ckpt.py  trainer ckpt -> vllm-loadable directory
tu/analysis/speed.py    time-to-τ / AUC / peak — THE dependent variable
tu/analysis/grad_probe.py  gradient geometry: eff_rank, rank_gain, cos_BC, cos_within
tu/analysis/final_figure.py, sft_curves.py, curves.py, catalysis.py, batch_reg.py
tu/tracing/             rollout trace features + rule-based span annotator

Trace data is gzipped

All rollout/trace .jsonl files are stored gzipped (16x smaller; 185MB -> 10MB). The loaders read .gz transparently — tu/offline/sft.py and tu/analysis/grad_probe.py accept either donor_x.jsonl or donor_x.jsonl.gz, and will fall back to the .gz if you pass the plain name. To materialize plain files instead:

gunzip -k transfer-unit/data/*.gz runs/*/rollouts.jsonl.gz

Reproducing in a new environment

1. Env: python 3.12, vllm==0.19.1, torch==2.10.0+cu128, transformers==5.13, fla-core, on CUDA-12 GPUs (H100 tested). Two GPUs per RL run (vLLM + policy), one GPU per SFT run.
2. Known environment traps (all handled in scripts/*.sh, read them before porting): fla-core's TileLang probe crashes → shim + FLA_TILELANG=0; fla refuses gated DeltaNet backward on Hopper w/ triton≥3.4 → monkeypatch to the pure-torch path; VLLM_ALLOW_INSECURE_SERIALIZATION=1 for callable collective_rpc; TORCH_BLAS_PREFER_CUBLASLT=1 for driver 535.
3. Paths are absolute (/mnt/bn/chobits-wx/jiayicheng/...) in scripts/ and a few defaults — grep for chobits-wx and repoint before running elsewhere.
4. Typical SFT-with-curve run: bash scripts/frozen_pipe_v2.sh <name> <B.jsonl> <C.jsonl> <c_select> <c_tokens> <seed> <gpu> "codegen:36"
5. Speed table over any run family: python -m tu.analysis.speed --runs-dir runs --pattern "mc_*_s*" --task codegen --taus 0.7

Headline numbers (B = codegen, tokens of B needed to reach pass@1 = 0.7)

mixed-in data	tt@0.7	speed-up
none (target data only)	15,215	1.0×
+90k more target data (single stream)	9,035	1.7×
+neutral format task 60k	6,305	2.4×
+easy math 60k	5,241	2.9×
+hard math 60k	3,435	4.4×
+own data 60k (mixed structure)	3,686	4.1×
+hard math 15k (sub-stoichiometric)	2,580	5.9×
+sql agent traces 60k	2,621	5.8×
+pyfix error-recovery 60k	2,105	7.2×

Full context, controls and caveats: reports/SPEED_THEORY.md.

Status and the honest caveat about the headline table

Active research snapshot (2026-08-29), negative-result-heavy — read this before using any number above:

• Round-2 predictions failed 1/5. The rank_gain criterion had no predictive power on five far-field domains (reports/PREREG_ROUND2.md): logic_grid was predicted fast and was the slowest additive; translate was predicted slow and was mid-pack.
• Total-token control removes most of the apparent speed-up. Mixing in C means more optimizer steps at the same count of B tokens. Re-expressed per total token, the right baseline is "just add more target data" (1.7x), not "add nothing" (1.0x), and most conditions fall to <=1.0x.
• Seed variance dominates. tt@0.7 spreads 6-7x across seeds within a condition (e.g. none: 3,330 / 20,806 / 21,508), comparable to or larger than between-condition differences. The headline table above is means over very noisy samples.
• A checkpoint-sampling bias was found: save_every counted optimizer steps, so high-C conditions were sampled more densely on the B-token axis, biasing them toward earlier crossings.

A preregistered high-seed decisive experiment (4 conditions x 8 seeds, eval n=48, B-token-equalized checkpoint spacing) addresses all three and was running at snapshot time: reports/PREREG_ROUND3.md, runs/r3_*. Treat the speed table as hypothesis, not result, until those land.