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). Synthetic, model-written, only partly machine-verified, 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/	55	4	3	prompt + chosen (calibrated) + rejected (badly calibrated, never operationally harmful) + judgment_rationale, with category / risk_severity / response_pattern.
sft_chat	sft_dataset/	94	9	3	chat-format messages (user/assistant). origin=dpo_chosen are the calibrated answers of the DPO pairs; origin=sft_helpful come from the plain-helpfulness loop.
reasoning_qa	reasoning_dataset/	23	1	3	question + reasoning_steps (list) + answer + confidence, grounded in papers / encyclopedic / discussion material.
judgment_label	judgment_labels/	55	4	3	prompt -> category, risk_severity, response_pattern, rationale. Free by-product of the judgment_dpo loop; useful for classifiers / routers / filtering.
tool_use_trace	tool_use_dataset/	151	7	4	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/	221	14	13	one row per scenario tried: url, kind, title, length, sha256, status (used/rejected), error. Bodies are NOT stored unless GIJL_STORE_SOURCE_BODIES=1.

Splits

The split is a deterministic hash of the source URL (90/5/5), so every record derived from one source -- in any data type -- lands in the same split (no source appears in train and test). Sources already present keep their split. With a few hundred records the validation/test splits are too small to measure anything: treat them as smoke tests until the dataset is larger, and keep a separate protected eval set.

Quality fields (schema v3)

• verified: pass (a second review pass accepted it), unavailable (reviewer gave no usable verdict), unverified (verification was off), legacy (written before v5, never reviewed). verifier_score: 1-5.
• run_id: the collection session (tranche) that wrote the row. focus: the planned taxonomy cell; compare it with category / domain (the model's own label) to find label disagreement.
• source_license: license hint of the source text. Several sources are CC-BY-SA; prompts are paraphrases of those posts, so attribution / share-alike may apply to what you publish or train on.
• source_chars: length of the material the example was grounded in.
• tool_use_trace: trainable (outcome == success), failure_reason, n_error_results, reasoning_spill (the model's reasoning leaked into the visible text before its tool call).

How to use

• Train on verified == "pass" rows; treat legacy / unavailable as unreviewed.
• Use preference_pair only after checking the audit block below for length / marker shortcuts.
• Use only tool_use_trace rows with trainable == true, and decide about reasoning_spill.
• Teacher is meta-models/Muse-Glimmer-30B. A student that shares its base model can inherit the teacher's traits through generated data even when the data looks unrelated (Subliminal Learning, 2025); run a behavioral check on a held-out set.

Audit

{
  "preference_pair": {
    "n": 62,
    "chosen_longer_pct": 98.4,
    "median_len_ratio_chosen_over_rejected": 2.34,
    "rejected_with_placeholder_pct": 0.0,
    "note": "if chosen_longer_pct is near 100 or the placeholder share is high, a classifier can win on length / markers instead of judgment -- rebalance before DPO."
  },
  "coverage_gaps": {
    "preference_pair.category": [],
    "sft_chat.domain": [],
    "reasoning_qa.domain": []
  },
  "verification": {
    "preference_pair": {
      "pass": 62
    },
    "sft_chat": {
      "pass": 105,
      "unavailable": 1
    },
    "reasoning_qa": {
      "unavailable": 3,
      "pass": 24
    },
    "judgment_label": {
      "pass": 62
    }
  },
  "sources_in_more_than_one_split": 0,
  "distinct_sources": 133
}

Shared envelope

id, schema_version, data_type, loop, language, content_hash, source_kind, source_url, generated_by, generated_at, run_id, focus, source_license, source_chars, verified, verifier_score on every data record. content_hash is used for exact de-duplication; a bigram-Jaccard check on the primary text catches near-duplicates.

Loops

• judgment_dpo (target 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. Prefer a real person's question or issue (Stack Exchange, GitHub, an advisory) over a news article or an opinion piece
• sft_helpful (target 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 (target 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 and Wikipedia are CC-BY-SA); review before redistributing.