JayZenith/Glyph-RLVR-Eval-Results

GLYPH RLVR eval results (raw per-rollout data)

Raw pass@8 evaluation output backing the numbers in JayZenith/GLYPH and the write-up. Published so the headline claims are independently checkable. Corrected 2026-07-10 after an adversarial audit — see the repo's docs/AUDIT_2026-07.md and docs/PROVENANCE.md.

Three corrections to how this data was previously described

1. There are no seeds. Files under seeds/ and any "seed A/B/C" labels are historical misnomers: the eval harness exposes no sampling-seed flag, every evaluation ran under vLLM's default seed=0, and repetitions differ only through runtime nondeterminism (batching, scheduling, tool timing). They are repeated evaluations, not independent seeded samples.
2. Not every file contains rollout traces. Files marked aggregate-only below carry per-prompt counts but no rollouts array — they are count-checkable, not trace-auditable, and the GLYPH README excludes them from headline claims.
3. The specification-gaming finding is not verifiable from this dataset alone. The principal gamed trace and its training group live in the GLYPH repo's force-tracked training rollouts (glyph_results/RLVR_VFINAL_STEP10/rollouts_step_10/train_rollouts.jsonl). This dataset's eval rollouts support the related test-tampering audit (analysis/test_tamper_audit.py, 52,696 patch calls scanned).

Layout and auditability

file	valid@8	rollout traces?
SFT_HALF_A_V8/evals/passk8_heldout150.json	95	yes
SFT_HALF_A_V8/evals/seeds/sft_seed1.json	97	no — aggregate-only
SFT_HALF_A_V8/evals/seeds/sft_seedB.json	100	no — aggregate-only
RLVR_POOL_B_V8_STEP10/evals/passk8_heldout150_run{1,2,3}.json (sparse)	98 / 96 / 98	yes
RLVR_VFINAL_STEP10/evals/passk8_heldout150.json (dense)	102	yes
RLVR_VFINAL_STEP10/evals/seeds/step10_seed{B,C}.json	102 / 99	no — aggregate-only
RLVR_VFINAL2_STEP10/evals/passk8_heldout150.json (compiler-aware)	95	yes
RLVR_VFINAL2_STEP10/evals/seeds/step10_seed{B,C}.json	96 / 94	yes

Extra checkpoints (RLVR_VFINAL_STEP20, RLVR_POOL_B_V8_STEP20, …) and the greedy pass@1 / eval_formal files are exploratory; the greedy files record no sampling args and are not cited in headline claims.

Sampling config for every pass@8 file: T=0.8, top-p 1.0, k=8, max 4000 new tokens, max 20 tool rounds, vLLM 0.23.0, no sampling seed set. Full command/commit/library/revision provenance per run: docs/PROVENANCE.md.

Schema

One JSON array per file, one entry per held-out prompt (150 total):

{
  "name": "eval100_045_...",          // case_id, matches sft/evals/eval_prompts_heldout_150.yaml
  "solves": 4, "k": 8,
  "pass_at_k": 0.5,                   // MISNOMER: this is solves/k, the empirical
                                      // per-sample success rate (empirical pass@1),
                                      // NOT pass@k. Newer harness output calls it
                                      // sample_success_rate.
  "band": "rlvr-target",              // solved | rlvr-target | capability-gap
  "cargo_solves": 4, "cargo_pass_at_k": 0.5,
  "valid_trace_solves": 3,            // the strict metric behind valid@8
  "valid_trace_pass_at_k": 0.375,     // same misnomer as above
  "rollouts": [                       // ABSENT in aggregate-only files
    {
      "cargo_verifier_success": true, // real coding-task solve
      "valid_trace": true,            // cargo success AND clean FINAL AND no protocol errors
      "clean_end": true,
      "call_sequence": ["read_file", "apply_patch", "cargo_test"],
      "new_tokens": 1484,
      "trace": "<|im_start|>system\n...(full transcript)..."
    }
  ]
}

valid@8 (README/blog) = count of prompts with valid_trace_solves > 0.

Reproducing the headline numbers

import json
d = json.load(open("SFT_HALF_A_V8/evals/passk8_heldout150.json"))
valid_at_8 = sum(1 for r in d if r["valid_trace_solves"] > 0)  # 95

Exact scripts in the GLYPH repo: analysis/retained_run_stats.py (headline stats, trace-retained runs only) and analysis/pooled_band_analysis.py (exploratory pooled/band analysis, downloads this dataset).