tts-sft/round4-sft-eval

SFT evaluation cut (2026-08-19)

The frozen held-out evaluation for the round-4 SFT arms (tts-sft/round4-sft), per SFT_DESIGN_2026-08-16.md §3/§7. Self-contained: test set + sampler + checker-aware grader + pass@k reporter. No training data here; the training corpora already exclude every id below.

Test set — testset.jsonl (657 problems, frozen seed 20260816)

set	n	purpose	metrics
dev (50 × buckets 1-3/4-7/8-11/12-15)	200	config/checkpoint selection + regression	pass@1, pass@16
h_frontier (80 only-SE / 20 only-ind / 20 both, bucket 0)	120	the 0→1 signal eval — touch ONCE with final configs	pass@1/8/16 + per-quadrant 0→1 (n=128 conditional, see protocol)
h_unsolved (bucket 0, nobody solved)	337	beyond-ceiling eval — final configs only	pass@1/8/16 + 0→1

Each row: wrapped generation prompt (byte-identical to arm B's), full official suite with time_limit=10, and for the 22 certified-SPJ problems the checker inline (grading is checker-aware by construction). testset_meta.json carries the sha256 and the frozen sampling protocol.

Running (shard per node — split however you like)

Serve the model (vLLM, OpenAI-compatible), then per node:

MODEL=<served name> OUT_DIR=eval_out/<model_tag> SHARD=0/4 K=16 \
  bash run_eval_shard.sh

SHARD=i/N deterministically takes every N-th problem of the sorted id list — any N works, nodes are independent, everything is resumable (rerun = continue). Alternatively pass an id file: eval_sample.py --ids my_shard.txt. When all shards of a model finish:

python3 eval_report.py --testset testset.jsonl \
    --grades "eval_out/<model_tag>/grades_*.jsonl" --ks 1,8,16 --out report.json

Protocol notes (please keep): sampling temp 1.0 / top_p 0.95 / max 16384 / reasoning medium / seed = 1234+k (already the defaults); dev may be re-run freely, the h_* sets are touch-once. Return eval_out/ (samples + grades + report) per model.

Comparison protocol (2026-08-20 amendment — supersedes "K=128 finalists"; pending Harman ack)

Every reported number is a three-way comparison at identical n: SE-SFT arm vs BoN-SFT control vs the UNTRAINED BASE model, all served and sampled under the identical protocol above. The base-at-same-n column is mandatory for any 0→1 claim — our own round diagnostics show reach grows with budget for any model, so a solve count without the base at the same budget is confounded with the budget itself.

Headline metrics are pass@1/8/16. Self-distillation's claim is amortization: the trained model should solve at small k what the base model needed a whole SE search to find. An effect visible only at large k means search was traded for brute sampling, not internalized.

n=128 is a conditional diagnostic, not a headline. Run it only if the n=16 result on the frontier sets is null or marginal, only on h_frontier (120 problems, ~0.2B tok/model), and always with base + control at the same n=128. Its two legitimate uses: (a) low-variance pass@1/8/16 — the reporter's unbiased estimator 1 − C(n−c,k)/C(n,k) uses all n draws; (b) distinguishing "no transfer" (p≈0) from "real but small transfer" (p≈1%), which look identical at n=16 but dictate different round-5 decisions. A positive n=128-only result is a mechanistic finding about the data, not evidence the model became useful at deployment budgets.

Validate after cloning (CPU only): TMPDIR=/var/tmp python3 validate_eval.py (12 tests: set integrity, stdin/SPJ/function-call grading paths, resume, pass@k math).