README.md

AGA-Code Stage B + Stage C Datasets

Procedurally generated, execution-verified training data for AGA-Code (aga_code.py), built under Guy's no-distillation rule: no teacher traces, no third-party data — every instance is a procedurally generated original with correctness established by running the code.

Files

File	Content
stage_b.jsonl	Deep-CoT SFT examples: instruction / think / critique / revise / answer / tests / family / difficulty / seed — consumed by stage_deepcot (<task>/<think>/<critique>/<revise>/<answer> sections)
stage_c.jsonl	Problem bank for stage_exec_grounded: instruction / tests / family / difficulty / seed / tier / mutation_score — no reference solutions (training file)
stage_c_verify.jsonl	Same rows as stage_c.jsonl plus reference_solution — held out for our own verification, NEVER fed to the model
manifest_v1.json	SHA-256s, counts, seed, gate results, per-family distribution, held-out family list, audit summary
BUILD_LOG.md	Per-family counts, discard counts and reasons, design decisions, known warts
audit_report_full.json	Independent audit re-execution results (pilot: audit_report_pilot.json)

Construction

• 60 program-generator families (string/list/dict ops, arithmetic, recursion, sorting/searching, number theory, parsing, simple OOP) with randomized parameters, 12 think skeletons, 7 grounded bug templates + clean-critique templates — all seeded with seed 1147 (MASTER_SEED).
• Stage B: each instance runs its reference + auto-generated tests; the answer is kept only if the tests PASS. ~80% of examples pair the reference with a real structurally-built bug (off-by-one, wrong operator, missing return, wrong variable, wrong edge, swapped branch, missing case) and a grounded critique/revise; ~20% are already-correct drafts whose critique explicitly says no bug was found and whose revise is a no-op.
• Stage C: problems are kept only if (a) the tests pass on the reference AND (b) at least one trivial-wrong solution fails the tests AND (c) a mutation score (≥5 mutants, up to 10, killed by the tests) ≥ 0.6. Instructions state the exact signature; examples are embedded worked asserts. Three tiers: easy (5,000 — single-function problems so an 11M model sees non-zero pass rates from the start), standard (3,500 — fresh instances from Stage B families, in-distribution), held_out (1,500 — whole families held out of Stage B, for transfer measurement).
• Held-out families (never in Stage B): caesar_cipher, run_length_encode, merge_sorted, prime_factors, collatz_steps, quicksort, csv_parse_line, eval_expr, bank_account, matrix_multiply.
• Dedup: SHA-256 on (instruction, answer) for Stage B, (instruction, tests) for Stage C; seed-1147 shuffle; no family > 10%, no instruction phrasing > 5%; no instruction overlap between Stage B and Stage C (checked on normalized text).
• Tokenizer note: the build performs no model generation or decoding, so no tokenizer is used; it is compatible with aga_code.py >= v0.1.7's fixed TokWrapper (ByteLevel decoder + skip_special_tokens=False).

Generator provenance

• Generator seed: 1147
• Generator script hashes (SHA-256): see manifest_v1.json (generator_sha256 — recorded as run)
• All 60 family names: see manifest_v1.json (all_families)

Scale

• Stage B: 25,000 examples (50 families; 10 held out). Difficulty: medium 12,680 / hard 7,546 / easy 4,774. Clean-critique fraction 20.4%.
• Stage C: 10,000 problems — easy 5,000 / standard 3,500 / held_out 1,500. Mutation score per problem ≥ 0.857 as shipped (gate ≥ 0.6); tier means 0.996 / 0.993 / 0.972.
• Gates: 100% of Stage B answers pass their tests; 100% of Stage C tests pass on references, reject a trivial-wrong solution, and kill ≥ 60% of mutants; B/C instruction overlap 0 after drop + top-up.
• Independent audit (audit.py): 50/50 shipped answers pass, 50/50 hand-broken versions fail — PASS. See BUILD_LOG.md for the full story, including the one known coverage wart (fib_recursive in Stage C).