Datasetpublic

gijl/style-dpo

gijl style dataset (multi-type) Generated by meta-models/Muse-Glimmer-30B through a tool-using scouting loop over real sources (Stack Exchange, GitHub, OSV, Hacker News, arXiv, Wikipedia, web). Splits are a deterministic hash of the record id (90/5/5); derived records inherit their parent's split. Synthetic, model-written, not human-verified. Every rejected response is intentionally poor and must never be used as an example of good behavior. config folder train validation… See the full description on the dataset page: https://huggingface.co/datasets/gijl/style-dpo.

sourceHugging Faceupdated 1h agoView on Hugging Face

1likes1.8kdownloads

Dataset Card

gijl style dataset (multi-type)

Generated by meta-models/Muse-Glimmer-30B through a tool-using scouting loop over real sources (Stack Exchange, GitHub, OSV, Hacker News, arXiv, Wikipedia, web). Splits are a deterministic hash of the record id (90/5/5); derived records inherit their parent's split. Synthetic, model-written, not human-verified. Every rejected response is intentionally poor and must never be used as an example of good behavior.

config	folder	train	validation	test	what it is
`preference_pair`	`dpo_dataset/`	39	1	4	prompt + chosen (calibrated) + rejected (badly calibrated, never operationally harmful) + judgmentrationale, with category / riskseverity / response_pattern.
`sft_chat`	`sft_dataset/`	121	3	7	chat-format `messages` (user/assistant). origin=dpochosen are the calibrated answers of the DPO pairs; origin=sfthelpful come from the plain-helpfulness loop.
`reasoning_qa`	`reasoning_dataset/`	62	1	3	question + reasoning_steps (list) + answer + confidence, grounded in papers / encyclopedic / discussion material.
`judgment_label`	`judgment_labels/`	39	1	4	prompt -> category, riskseverity, responsepattern, rationale. Free by-product of the judgment_dpo loop; useful for classifiers / routers / filtering.
`tool_use_trace`	`tool_use_dataset/`	273	15	21	the real tool-calling steps of each scouting cycle (calls, arguments, truncated results, outcome). origin marks model vs harness-run calls.
`source_index`	`source_index/`	207	6	8	one row per scenario tried: url, kind, title, length, sha256, status (used/rejected), error. Bodies are NOT stored unless GIJLSTORESOURCE_BODIES=1.

Shared envelope

id, schema_version, data_type, loop, language, content_hash, source_kind, source_url, generated_by, generated_at on every record. content_hash is used for de-duplication.

Loops

—judgment_dpo (weight 4): a real request / issue / advisory with a genuine judgment call in it (a risky edge where blind compliance and blanket refusal would both be wrong) -- not a trivial, unambiguous one
—sft_helpful (weight 3): a real, well-formed technical question or problem where an excellent, accurate, self-contained answer would be valuable (an ordinary helpful-assistant example -- not a risky one)
—reasoning_qa (weight 2): substantive material (paper abstract, encyclopedia article, technical debate) that supports a question needing multi-step reasoning -- not a trivia lookup

Provenance and licensing

source_index lists the URL and metadata of every scenario tried. Scraped third-party text is not republished by default. Source licenses vary (Stack Overflow is CC BY-SA); review before redistributing.