Update Dataset Card for spec-verify

README.md

@@ -1,32 +1,66 @@
 ---
-dataset_info:
-  features:
-  - name: id
-    dtype: string
-  - name: source
-    dtype: string
-  - name: repo
-    dtype: string
-  - name: specification
-    dtype: string
-  - name: code
-    dtype: string
-  - name: url
-    dtype: string
-  splits:
-  - name: train
-    num_bytes: 35613638
-    num_examples: 18597
-  - name: test
-    num_bytes: 3958347
-    num_examples: 2067
-  download_size: 18097126
-  dataset_size: 39571985
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
-  - split: test
-    path: data/test-*
+license: mit
+language:
+- en
+pretty_name: spec-verify v2 (English, quality-filtered spec + code)
+tags:
+- software-engineering
+- specification
+- code
+- github
+- english
+- quality-filtered
+size_categories:
+- 10K<n<100K
 ---
+
+# spec-verify dataset **v2**
+
+## Description
+
+This is the **v2** dataset on the Hugging Face Hub (`spec-verify-dataset-v2`), built for **10,000+** high-quality **(specification, code)** pairs.
+
+- **Quality filtered**: non-English and low-signal rows are removed using heuristics (ASCII/Latin ratio, repetition, minimum word counts, valid docstring shape for Python rows).
+- **English-oriented**: specifications are cleaned to **ASCII/Latin**-heavy text; issue/PR and docstring rows must pass English dominance checks.
+- **Real code for docstring rows**: the `docstring` subset requires **substantial Python implementation** (not signature-only), including logic keywords and length checks.
+- **GitHub issues / PRs**: natural-language specifications only (`code` may be empty); filtered for English and non-repetitive text.
+
+Pairs are mined from public GitHub repositories. Intended for **spec-to-code**, **code-to-spec**, and **verification** research.
+
+## Data fields
+
+| Column | Meaning |
+|--------|---------|
+| `id` | Stable hash identifier for the row |
+| `source` | `github_issue` (includes PR descriptions in this schema) or `docstring` |
+| `repo` | Source repository in `owner/name` form |
+| `specification` | English-cleaned NL spec (issue/PR text, or signature + docstring + optional structured notes from extraction) |
+| `code` | Full function implementation for `docstring` rows; often empty for `github_issue` |
+| `url` | Link to the issue, pull request, or file on GitHub |
+
+## Example record
+
+```json
+{
+  "id": "d3b066c9609ab8e6",
+  "source": "github_issue",
+  "repo": "aio-libs/aiohttp",
+  "specification": "GunicornWebWorker fails to reset SIGCHLD causing spurious \"Worker exited\" errors\n\n### Describe the bug\n\nThe `aiohttp.worker.GunicornWebWorker` class overrides `init_signals()` but fails to reset SIGCHLD to SIG_DFL. This causes workers to inherit the gunicorn master arbiter's SIGCHLD handler, leading to spurious error logs when application code spawns subprocesses.\n\n## Evidence\n\n### 1. aiohttp/worker.py has incomplete fix\nLines 191-192 contain a comment about resetting signals, but NO implementation:\n\n```python\n# Reset signals so Gunicorn doesn't swallow subprocess return codes\n# See: https://github.com/aio-libs/aiohttp/issues/6130\n#   CODE IS MISSING HERE!\n```\n\n### 2. Base Worker resets SIGCHLD correctly\n`gunicorn/workers/base.py` properly resets SIGCHLD:\n\n```python\nSIGNALS = [\"ABRT\", \"HUP\", \"QUIT\", \"INT\", \"TERM\", \"USR1\", \"USR2\", \"WINCH\", \"CHLD\"]\n\ndef init_signals(self):\n    # reset signaling\n    for s in self.SIGNALS:\n        signal.signal(s, signal.SIG_DFL)  #   Resets SIGCHLD\n```\n\n### 3. aiohttp Worker doesn't call super()\n`aiohttp/worker.py:160-193` overrides `init_signals()` but never calls `super().init_signals()`, so SIGCHLD is never reset.\n\n### To Reproduce\n\n1.Install gunicorn and aiohttp latest versions\n2.Run this code from below, see usage inside\n```\n#!/usr/bin/env python3\n\"\"\"\nReproduction of aiohttp GunicornWebWorker SIGCHLD inheritance bug.\n\nUsage:\n    python aiohttp_sigchld_bug.py          # Show the bug\n    python aiohttp_sigchld_bug.py --fixed  # Show with fix\n\nBug: Workers inherit master's SIGCHLD handler, logging spurious errors\nabout subprocess PIDs as if they were worker PIDs.\n\nExpected: Clean worker startup\nActual (without fix): ERROR logs like \"Worker (pid:X) exited with code 1\"\nwhere X never appears in \"Booting worker\" logs.\n\"\"\"\nimport sys\n\n# Run with gunicorn when executed directly\nif __name__ == \"__main__\":\n    import os\n    import subprocess\n    import tempfile\n\n    apply_fix = \"--fixed\" in sys.argv\n\n    # Create a separate module that spawns subprocesses during import\n    submodule_content = \"\"\"\nimport subprocess\nimport time\n\n# Spawn subprocesses during module import to trigger SIGCHLD\n# This simulates real-world code that checks versions, calls git, dpkg, etc.\nprocesses = []\nfor i in range(5):\n    # Spawn subprocess that exits with code 1\n    p = subprocess.Popen(['false'], stdout=subprocess.PIPE, stderr=subprocess.PIPE)\n    processes.append(p)\n\n# Small delay creates window where:\n# 1. Subprocesses exit and become zombies\n# 2. SIGCHLD is delivered to worker\n# 3. If worker has inherited arbiter's handler, it reaps OTHER workers' zombies\ntime.sleep(0.05)\n\n# Clean up our zombies\nfor p in processes:\n    p.wait()\n\"\"\"\n\n    # Create the main app module\n    app_content = \"\"\"\n# Import submodule that spawns subprocesses (triggers bug)\nimport subprocess_module\n\nfrom aiohttp import web\n\nasync def hello(request):\n    return web.Response(text=\"Hello World\\\\n\")\n\nasync def application():\n    app = web.Application()\n    app.router.add_get(\"/\", hello)\n    return app\n\"\"\"\n\n    # Create config with optional fix\n    config_content = \"\"\"\nworkers = 4\nworker_class = \"aiohttp.worker.GunicornWebWorker\"\nbind = \"127.0.0.1:8765\"\nloglevel = \"info\"\n\"\"\"\n\n    if appl
+… (truncated for card)
+```
+
+## Intended use
+
+- Training and evaluating models that **generate formal or behavioral specs from natural language**
+- **Spec-to-code** or **code-to-spec** alignment with **implementation-heavy** Python for docstring-sourced rows
+- **Test / verification** research conditioned on specifications
+
+## Limitations
+
+- Sources are public GitHub data; quality and licensing follow upstream repositories.
+- Heuristic filters favor English and Latin script; they do not guarantee semantic correctness.
+- Older **v1** uploads may use a different schema; this repo is **v2** (`spec-verify-dataset-v2`).
+
+## Citation
+
+If you use this dataset, cite the upstream repositories and acknowledge the spec-verify extraction pipeline.