Datasets:

formalanon
/

semantic-lean-errors

@@ -1,194 +1,11 @@
-# semantic_lean_errors
-`semantic_lean_errors` is a release bundle for semantic error auditing of NL → Lean 4 formalizations.
-It packages a 92-problem human-labeled benchmark together with the prompt, schema, and taxonomy used for the paper's 6-category semantic-audit experiments.
-## Directory Overview
-This directory is meant to be self-contained for release.
-It includes:
-- `semantic_lean_errors.jsonl`
-  - Canonical machine-facing dataset.
-  - One record per line.
-  - Recommended file for Hugging Face and downstream use.
-- `semantic_lean_errors.json`
-  - Readable companion file.
-  - Stores the same records as `semantic_lean_errors.jsonl`, but as a pretty-printed JSON array.
-- `ground_truth_schema.md`
-  - Schema for the exported records in this release bundle.
-- `error_tags.md`
-  - Taxonomy definitions, boundary rules, and detail-tag guidance.
-- `eval_prompts_v4_fewshot.md`
-  - Few-shot prompt set used for the category-specific semantic-audit evaluation.
-- `README.md`
-  - Dataset card and release notes.
-## Dataset Composition
-The benchmark combines four source datasets:
-- FormalMath: `22`
-- ProofNet: `13`
-- ProverBench: `23`
-- CombiBench: `34`
-Total:
-- `92` labeled problems
-- `87` error examples
-- `5` clean examples
-Primary-category counts:
-- `problem_statement_error`: `5`
-- `specification_error`: `24`
-- `formalization_error`: `8`
-- `domain_mismatch`: `14`
-- `definition_mismatch`: `26`
-- `quantifier_indexing_mismatch`: `10`
-- clean examples (`error_primary = null`): `5`
-Additional label structure:
-- rows with non-empty `error_secondary`: `33`
-- rows with non-empty `error_tags`: `87`
-## Record Schema
-Each row is a single NL/Lean pair with human labels.
-The exported per-record fields are:
-- `id`
-- `source`
-- `problem_nl`
-- `lean_code`
-- `expected_verdict`
-- `error_primary`
-- `error_secondary`
-- `error_tags`
-- `meta_tags`
-- `gold_explanation`
-Field meanings are documented in `ground_truth_schema.md`.
-Two label views are represented:
-1. Binary semantic-error detection
-- Use `expected_verdict`.
-- `YES` means the row contains some semantic error.
-- `NO` means the row is a clean example.
-2. Six-category semantic audit
-- Use `error_primary`.
-- In the current evaluation setup, a queried category expects `YES` iff `error_primary` equals that category, and `NO` otherwise.
-## Prompt Overview
-The included prompt file `eval_prompts_v4_fewshot.md` defines a 6-category evaluation setup.
-For each NL/Lean pair, the model is queried once per category:
-- `problem_statement_error`
-- `specification_error`
-- `formalization_error`
-- `domain_mismatch`
-- `definition_mismatch`
-- `quantifier_indexing_mismatch`
-Each prompt asks the model to judge one specific category and return structured JSON with:
-- `Verdict`
-- `Explanation`
-- `DetailTags`
-- `NeedsReview`
-Because the prompt is category-specific, the 92 labeled problems expand to `552 = 92 × 6` category-level classifications in the paper's evaluation setup.
-## Experimental Reproducibility
-This release includes the exact benchmark and prompt family used for the paper's semantic-audit experiments.
-Evaluation setup:
-- benchmark: `92` labeled problems
-- categories: `6`
-- total category-level judgments: `552`
-- prompt file: `eval_prompts_v4_fewshot.md`
-The paper reports results for at least these model configurations on this benchmark:
-- Sonnet 4.5 + Thinking
-- GPT-5.2 + Thinking
-### Paper-Reported Aggregate Results
-These are the paper-reported numbers for the 92-problem semantic audit and should be treated as the canonical release numbers for this version of the benchmark.
-| Model | Prec. | Rec. | F1 | Acc. | Cost |
-|-------|------:|-----:|---:|-----:|-----:|
-| Sonnet 4.5 + Thinking | 0.30 | 0.82 | 0.42 | 68.1% | $13.71 |
-| GPT-5.2 + Thinking | 0.24 | 0.91 | 0.37 | 54.6% | $6.86 |
-### Paper-Reported Per-Type F1
-| Error Type | Sonnet 4.5 | GPT-5.2 |
-|-----------|-----------:|--------:|
-| Problem Statement Error | 0.29 | 0.10 |
-| Specification Error | 0.51 | 0.48 |
-| Formalization Error | 0.18 | 0.15 |
-| Domain Mismatch | 0.33 | 0.27 |
-| Definition Mismatch | 0.57 | 0.54 |
-| Quantifier/Indexing | 0.39 | 0.33 |
-Interpretation:
-- both models have relatively high recall and lower precision
-- both models perform best on `specification_error` and `definition_mismatch`
-- both models struggle most on `formalization_error`
-Note on reproducibility:
-- this directory contains the benchmark, prompt, schema, and taxonomy
-- if the exact evaluation pipeline is rerun later, small differences can arise from model/provider/version changes
-- the table above reflects the paper-reported results for this release
-## Normalization Policy
-This release intentionally uses a single canonical code field:
-- `lean_code` is the only exported statement field
-- ProofNet rows are normalized from `original_lean_code` to `lean_code`
-- `fixed_lean_code` and `lean_code_fixed` are omitted from this canonical release
-- list-valued annotation fields are normalized to lists
-This avoids ambiguous rows that contain two different Lean statements under one set of labels.
-If repaired variants are needed for comparison studies, they should live in a separate paired comparison artifact.
-## Taxonomy
-The category taxonomy and detail-tag definitions are included in `error_tags.md`.
-That file defines:
-- the six primary semantic-error categories
-- boundary rules between overlapping categories
-- the allowed detail-tag families
-- optional meta-tag conventions
-## Release Cleanup Applied
-Before export, four internally inconsistent rows were corrected from `expected_verdict = YES` to `expected_verdict = NO` because they were clean examples with `error_primary = null` and explanations indicating no semantic error:
-- `formalmath_olymid-ref-base_5274`
-- `proofnet_rudin_exercise_2_28`
-- `proofnet_rudin_exercise_3_21`
-- `proofnet_axler_exercise_6_16`
-After this cleanup, the release contains `87` error rows and `5` clean rows.
-## Recommended Use
-For release and archival use:
-- treat `semantic_lean_errors.jsonl` as the canonical dataset file
-- keep `README.md`, `ground_truth_schema.md`, `error_tags.md`, and `eval_prompts_v4_fewshot.md` alongside it
-- treat `semantic_lean_errors.json` as a readable companion artifact rather than the canonical training/eval file

+---
+configs:
+  - config_name: default
+    data_files:
+      - split: train
+        path: semantic_lean_errors.jsonl
+---
+# Semantic Lean Errors
+92 expert-annotated examples of semantic errors in autoformalized Lean 4 mathematics.