zli12321/gos-skill-retrieval-embeddings

Graph of Skills retrieval vectors

Fixed text-embedding-3-large vectors (3,072 dimensions, float32) for the skill libraries and task queries used by the OpenRSI task rsi/gos-skill-retrieval. They let Graph of Skills default retrieval run offline: the task's starting retriever seeds personalized PageRank from these vectors and lexical candidates, exactly as GoS's engine does.

File	Contents
skills_1000.npy, skills_2000.npy	One row per skill, in the order of manifest.json libraries.<library>.rows (skill directory names).
dev_query_embeddings.npy	One row per public dev task, in the order of dev_query_embedding_ids.json.
test_query_embeddings.npy	56 hidden-query vectors for task preparation, in ascending task-ID order within the fixed test split. SHA-256: 1f57eb78e3317b8e10e5d0211eaad226d7b74148899c934590db07d163dc3513.
manifest.json	Model and dimension, plus text and file SHA-256 values for the skill and dev vectors. The test-vector checksum is recorded above.

How the vectors were made

• Skills: the skills_1000 and skills_2000 libraries from DLPenn/graph-of-skills-data at 83a8c59f6d09b849e57a50ed2577e253995d257e, parsed with GoS's parse_skill_document (900-character snippets) at GoS commit 203f60a2c689da055ce1ac351eb3cb9912a3bca7. Each row embeds GoS's node text, SkillNode.from_lists(...).to_str().
• Queries: SkillsBench task instructions from benchflow-ai/skillsbench at 9a1f4dd5f7659f75707435da3ce854b6e48321d1. Each row embeds GoS's query text without LLM rewriting, _fallback_query_schema(instruction).to_query_text(), which is GoS's default (ENABLE_QUERY_REWRITE is off).
• Model: text-embedding-3-large through an Azure OpenAI deployment, computed once in October 2026. Vectors are stored as returned, without renormalization.

The hidden-query vector file contains no task IDs, instructions, required-skill labels, or per-task outcomes. Its rows follow the fixed, stratified 56-task split of the 87 pinned SkillsBench tasks, sorted by task ID. Task preparation must place the file in Judge-only tests, not in the agent's Work image. Publishing these vectors reduces secrecy of the test inputs, so evaluation must still withhold test instructions, labels, and per-task feedback from Work and apply the task's anti-hard-coding rules.

Licenses

The vectors are released under MIT. They are derived from the MIT-licensed GoS skill libraries and from SkillsBench instructions (Apache-2.0).