Datasets:

EricSpencer00
/

tla-prover-results

Error code:   FeaturesError
Exception:    ArrowInvalid
Message:      Schema at index 1 was different: 
run_id: string
control_id: string
statement: bool
timeout_seconds: int64
result: string
elapsed_seconds: double
experiment_type: string
vs
arm: string
model: string
framing: string
prompt_version: string
specs_n: int64
pass_at_32_successes: int64
pass_at_32_trials: int64
sample_passes: int64
sample_total: int64
sample_rate: double
wilson_95_ci_lower: double
wilson_95_ci_upper: double
prompt_comparable_to_baseline: bool
notes: string
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/split/first_rows.py", line 244, in compute_first_rows_from_streaming_response
                  iterable_dataset = iterable_dataset._resolve_features()
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 4408, in _resolve_features
                  features = _infer_features_from_batch(self.with_format(None)._head())
                                                        ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2679, in _head
                  return next(iter(self.iter(batch_size=n)))
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2861, in iter
                  for key, pa_table in ex_iterable.iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2395, in _iter_arrow
                  yield from self.ex_iterable._iter_arrow()
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 580, in _iter_arrow
                  yield new_key, pa.Table.from_batches(chunks_buffer)
                                 ~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^
                File "pyarrow/table.pxi", line 5040, in pyarrow.lib.Table.from_batches
                File "pyarrow/error.pxi", line 155, in pyarrow.lib.pyarrow_internal_check_status
                  return check_status(status)
                File "pyarrow/error.pxi", line 92, in pyarrow.lib.check_status
                  raise convert_status(status)
              pyarrow.lib.ArrowInvalid: Schema at index 1 was different: 
              run_id: string
              control_id: string
              statement: bool
              timeout_seconds: int64
              result: string
              elapsed_seconds: double
              experiment_type: string
              vs
              arm: string
              model: string
              framing: string
              prompt_version: string
              specs_n: int64
              pass_at_32_successes: int64
              pass_at_32_trials: int64
              sample_passes: int64
              sample_total: int64
              sample_rate: double
              wilson_95_ci_lower: double
              wilson_95_ci_upper: double
              prompt_comparable_to_baseline: bool
              notes: string

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

TLA-Prover Aggregate Evaluation Results

This release contains compact, provenance-linked aggregate results from the TLA-Prover project. It combines a completed July 2026 W4 evaluation with a September 2026 checker startup diagnostic. These are different experiment types and must not be interpreted as one benchmark.

The release contains no holdout specifications, per-spec rows, prompts, model outputs, checkpoints, training data, or machine-local paths.

Files

generation_metrics.csv: four measured framing-A arms, with sample counts, Wilson confidence intervals, and pass@32 counts.
paired_comparisons.csv: the two-level paired-bootstrap comparisons, including the same-model control.
control_diagnostic.csv: three TRUE/FALSE checker controls from the latest completed September diagnostic. These are checker controls, not model results.
metadata.json: source revisions, source-file hashes, definitions, limitations, and explicit exclusions.

Main Finding

W4-diamond-gold produced 105/960 passing samples (10.94%) versus 47/963 (4.88%) for v2 SFT on the 30-spec framing-A evaluation. The paired difference was +0.0604 (6.04 percentage points; 95% CI [+0.0208, +0.1031], p=0.0019). The comparison with the untuned base was +0.0406 (4.06 points; 95% CI [-0.0021, +0.0854], p=0.0641), which does not establish a statistically significant gain at the conventional 0.05 level.

The post-prompt-fix W4 result is 130/960 samples and 16/30 pass@32, but 13 of 30 prompts changed. It is therefore reported separately and is not treated as a matched comparison with the earlier arms. Gate 2 was not met.

Method

The July aggregate analysis re-scores existing run ledgers without new inference. It uses per-sample pass rate with 95% Wilson intervals, plus a two-level paired bootstrap that resamples specs and rows within each spec. Cross-arm comparisons include only specs with matching prompt_sha256. The control comparison uses the same model and prompts across two runs and is expected to be consistent with a null difference.

Reproduce the July analysis in the source repository with:

python3 tools/rowlevel_power.py

The September diagnostic checks checker startup and negative-control rejection only. Its three classifications are two proved_control outcomes and one expected negative rejection. It had zero candidate checks, zero protected checks, and zero training updates. It provides no model, proof-quality, or gate credit.

Limitations

The W4 evaluation uses a frozen 30-spec holdout; aggregate rows are provided without the underlying examples or prompts.
pass@32 is a coarse per-spec summary and should be read with the per-sample rates and paired analysis.
The W4-versus-untuned-base comparison is inconclusive at p=0.0641.
The 16/30 post-prompt-fix result is not directly comparable because 13 prompts changed.
No result in this release establishes that Gate 2 was passed or that a model can produce verified TLA+ proofs.
The September startup diagnostic is operational evidence, not a benchmark score.

Provenance and License

The July metrics are transcribed from results/analysis/rowlevel_reanalysis_2026-07-30.md at source commit 629f2b1ca846638cce2f0631d7fb5248c92aa87c. The September control summary is derived from results/runs/protected-pilot-resume-20260922-v1/macmini-verification-v2/remote-control-startup-v1/summary.json at source commit 2d7728799db30e60f95dd48db0834967857070d2. Source SHA-256 values are recorded in metadata.json.

This aggregate release is provided under MIT, matching the project repository license. No source specifications, upstream datasets, or third-party software are redistributed here; their respective licenses remain separate.

Downloads last month: 52

Total file size:

10.6 kB