Denniscp8222's picture
Re-sync to the current corpus index (3,454 blocks): drop the redacted placeholder entry from the first build; README counts corrected
29de092 verified
|
Raw History Blame Contribute Delete
1.8 kB
metadata
license: mit
task_categories:
  - text-generation
  - fill-mask
language:
  - en
tags:
  - code
  - multi-agent
  - provenance
  - deduplicated
pretty_name: ASIN-HHC Code Corpus
size_categories:
  - n<1K
  - 1K<n<10K

ASIN-HHC Code Corpus

3,454 unique code blocks extracted from 13 months of builds across six AI platforms (DeepSeek, Gemini, Grok, ChatGPT, Kimi, CodePen), deduplicated by SHA-256 of normalized content, each with provenance back to its source conversation.

What it is

Every fenced code block from 1,817 conversation files, plus CodePen pen sources. Deduplicated (5,454 raw → 3,454 unique). Secrets, API keys, emails, and phone numbers were detected and excluded (371 blocks).

Re-synced 2026-11-18 to the current corpus index: one redacted placeholder entry from the first build was dropped, bringing the set to 3,454 blocks.

What it is NOT

Build material, not verified evidence. Inclusion means "an agent wrote or pasted this code," not "this code was tested or reviewed." Provenance is tracked so any block can be traced to its source; nothing is promoted beyond what its source supports.

Fields

  • sha256 — full SHA-256 of normalized block content
  • language — detected language (python, javascript, html, bash, json, …)
  • purpose — keyword-classified purpose tag (receipt, gate, wallet, …)
  • bytes, occurrences — size and raw occurrence count
  • mirrors — which of the six AI platforms the block appeared in
  • source_count — number of provenance records
  • code — the code block text

Source

Built by the ASIN-HHC programme. Extractor: tools/build_corpus.py (deterministic, mirrors read-only). Full repo with per-block files and index: https://github.com/dbottrader/ASIN-HHC-Code-Corpus

MIT licensed. Equality before profits.