PRD — Micro-AutoMixer for specialist pretrain ratios

Goal

Replace the current “hand-wave a 50/30/20 code/web/math split” pattern with a small AutoMixer-style ratio search. Train ~6–12 short proxy runs across candidate ratios, score them on a fixed capability suite (BFCL + GSM8K + HumanEval + Pace unhappy-paths), fit a simple quadratic surrogate, propose the next mix. Stop when proposed gains fall below a threshold.

Scaled down from Poolside’s Laguna recipe (Laguna deep dive) to a single Mac.

Why now

• A1 specialist training is the upcoming work; current ratios are guesses. Even a 3% absolute improvement on BFCL from a better mix is bigger than most of our other Tier-B knobs.
• Each proxy run is small (~30 min on a 22M model, ~3 hours on the flagship-class). The full search is ~12 × 30 min = 6 hours wall- clock for the small-proxy regime — overnight runs return a ratio recipe.
• The eval surfaces all exist (BFCL, τ-bench wired in E1/E2). The search loop and the surrogate fit are the new pieces.

Scope — in

• Sources/TinyGPT/AutoMix.swift — orchestrator. Inputs:
  • A list of corpora (--corpus name1=path1.txt --corpus name2=path2.txt ...) — these are the variables to ratio.
  • A list of eval tasks (--task bfcl --task gsm8k ...) — uses the existing tinygpt run-lm-eval + tinygpt eval-bfcl paths.
  • Search budget: --proxy-runs N --proxy-steps S (e.g. 8 × 2000 steps).
• Search loop:
  1. Sample N initial mixes (Dirichlet over corpora) including the “uniform” anchor.
  2. Train each proxy run.
  3. Score on the tasks.
  4. Fit a quadratic surrogate on (ratio, score).
  5. Propose the next mix to evaluate (Expected Improvement over the current best).
  6. Stop when EI < threshold or budget exhausted.
• Output: automix-report.jsonl with one row per proxy run + a final automix-recommendation.json carrying the best ratio + predicted full-scale lift.

Scope — out

• Full-scale AutoMix (replacing the small proxy with a near-full run). Defer — proxy-to-full transfer is the open research question. V1 uses small proxies and trusts the small/full correlation; documented as a caveat.
• Multi-objective fronts (Pareto across speed vs accuracy). V1 uses a fixed weighted sum.
• In-the-loop AutoMix (mix changes during a single run). Way bigger scope; defer.

Files to touch

Don’t touch

• tinygpt train itself — automix orchestrates tinygpt train subprocesses. No coupling.

Acceptance criteria

• tinygpt automix --corpus code=stack.txt --corpus math=meta.txt --corpus web=fineweb.txt --task bfcl --proxy-runs 6 --proxy-steps 2000 --out auto.jsonl runs end-to-end on the M5 Pro.
• The surrogate’s EI estimate at step N+1 correlates positively with the actual measured improvement (a sanity check that the fit isn’t degenerate).
• automix-smoke.sh passes in CI on tiny corpora.
• docs/recipes/automix.md walks through interpreting an automix-report.jsonl.

Reference patterns

• Sources/TinyGPT/RunLmEval.swift — the post-checkpoint subprocess-orchestration pattern. Same shape: spawn child, parse JSONL, aggregate.
• Poolside Laguna deep dive — the automixing-before-pretrain motivation. Cite, don’t redocument.
• DoReMi, Xie et al. 2023 — the formal mixture-optimization treatment. Our V1 uses the simpler EI/ quadratic surrogate; cite as the rigor target for V2.

Open questions

• Whether to use Bayesian optimization (Gaussian process) over the quadratic surrogate for the EI step. Recommendation: quadratic for V1 — 200 LOC; GP is 500+ and a dependency.
• Proxy-to-full transfer assumption. Recommendation: document as a caveat; cite Poolside’s claim that 10x-smaller proxies retain rank-correlation on capability evals.