SafetyMP/corporate-site-harness-training-data

Dataset Card for Corporate Site Harness Training Data

Revision: v0.3-lora-standard

• Factory git commit: ad72ba654590dd83d7070b23990b164ffae688a4

Dataset Summary

English chat-style supervised fine-tuning (SFT), preference (DPO), and held-out evaluation data for teaching a local LLM the corporate/site harness used by corporate-site-harness:

• policy — phases, roles, workspace isolation, premium-model routing, factory vs product
• cli — corp-harness argv, tool-grounded status/route JSON, multi-turn trajectories
• artifacts — gate reports, evidence, handoff digests, bad-report repair
• refusals — stratified hard negatives (actor-user, invent-PASS, nested roots, …)

Rows are chat messages (single-turn or multi-turn trajectory) suitable for Unsloth, TRL SFTTrainer, and other ChatML / messages loaders.

This dataset is a training/eval resource only. It is not a harness gate and does not grant program approval. A green fine-tuned model does not mean a harness PASS.

Intended Use

In scope

• Specialize a chat model (recommended: Qwen2.5-7B-Instruct via QLoRA) toward harness-safe explanations, CLI composition, artifact shapes, and refusals.
• Offline evaluation with score_eval.py on the validation split.
• Optional DPO on the dpo config (refusal-derived preference pairs).

Out of scope

• Replacing corp-harness executable gates, digests, or user approval.
• Claiming production “governance” compliance solely from model generations.
• Training on the validation split.
• Using rejected DPO completions as positive SFT targets.

Dataset Structure

Configurations

Config	Files	Purpose
default	data/train.jsonl, data/validation.jsonl (+ Parquet twins)	Full SFT rows
messages_only	data/messages_only_*.jsonl (+ Parquet)	Slim messages only
dpo	data/dpo_train.jsonl, data/dpo_validation.jsonl (+ Parquet)	Preference pairs

Parquet twins use the same logical columns with nested fields JSON-serialized for Hub viewer portability (messages, conversations, prompt, etc.).

Data Instances

from datasets import load_dataset

ds = load_dataset("SafetyMP/corporate-site-harness-training-data")
example = ds["train"][0]
print(example["id"], example["domain"])
print(example["messages"])

dpo = load_dataset("SafetyMP/corporate-site-harness-training-data", "dpo")
print(dpo["train"][0]["chosen"][:120])

Data Fields (default)

Field	Type	Description
id	string	Stable unique id
domain	string	policy, cli, artifacts, or refusals
task_type	string	qa, decision, cli_compose, json_emit, or refuse
difficulty	string	easy, medium, or hard
messages	list of {role, content}	OpenAI-style chat turns
conversations	list of {from, value}	ShareGPT-style turns
source_refs	list[string]	Doc / policy references
tags	list[string]	Topic tags
grading_json	string	Eval rubric JSON (validation); empty on train

Data Fields (dpo)

Field	Type	Description
id	string	Preference pair id
domain	string	Usually refusals
prompt	list of messages	Context up to the final assistant turn
chosen	string	Gold refusal / safe completion
rejected	string	Synthetic unsafe compliance (no real secrets)

Data Splits

Split	default rows	dpo rows
train	3000	400
validation	450	60

Train and validation dialogues are disjoint (enforced by validate_dataset.py).

Training Recipe (Qwen2.5-7B QLoRA)

Endorsed recipe (user-run; not executed by maintainers in the packaging pass): datasets/harness-llm/recipes/train_qlora_qwen25_7b.py

Knob	Value
Base	Qwen/Qwen2.5-7B-Instruct
Method	QLoRA (Unsloth or TRL SFTTrainer)
Rank / alpha	r=16, lora_alpha=16
Learning rate	≈ 2e-4
Epochs	1–2
Max sequence length	4096
Loss masking	train_on_responses_only (assistant turns)
Data	Hub default or local datasets/harness-llm/huggingface/

Do not train on validation. Optional preference tuning: load config dpo after SFT.

Evaluation Protocol

1. Keep validation held out (never in the SFT trainer train split).
2. Generate assistant replies for validation prompts; write a JSONL of {"id", "messages"} (or reuse gold structure with model assistant text).
3. From the factory checkout, run:

python3 datasets/harness-llm/scripts/score_eval.py \
  --predictions path/to/model_generations.jsonl

4. Report overall and by-domain scores (policy, cli, artifacts, refusals).
5. Interpret results as model specialization metrics only. A high score is not a harness gate PASS and does not grant user approval.

Gold self-score on the packaged eval set must be 100% via the same scorer (CI / local verification).

Dataset Creation

Curation Rationale

Agents frequently invent harness shortcuts (--actor user, forged PASS, nested roots). This corpus teaches correct CLI/policy language and hard refusals without replacing executable gates.

Source Data

Synthetic, expert-authored templates expanded in datasets/harness-llm/scripts/ against corporate-site-harness docs and CLI contracts. Paths are workspace-neutral (/work/...); no personal data.

Annotations

Authoritative assistant targets and eval grading rubrics are curated with the dataset scripts. DPO rejected strings are synthetic unsafe paraphrases, not logged user traffic.

Personal and Sensitive Information

No PII, credentials, or live secrets. Placeholder digests only.

Considerations for Using the Data

Social Impact of Dataset

Intended to reduce unsafe harness shortcut suggestions in specialized assistants. Misuse: treating generations as governance authority.

Discussion of Biases

Harness-specific, English-only, refusal-heavy relative to open-ended coding.

Other Known Limitations

• Mid-small scale (1K<n<10K).
• Heuristic grading, not a full semantic judge.
• Green model ≠ harness PASS.

Additional Information

Dataset Curators

SafetyMP — corporate-site-harness maintainers.

Licensing Information

Apache License 2.0.

Citation Information

@misc{safetymp_corporate_site_harness_llm,
  title        = {Corporate Site Harness Training Data},
  author       = {SafetyMP},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/SafetyMP/corporate-site-harness-training-data}},
  note         = {Revision v0.3-lora-standard; derived from https://github.com/SafetyMP/corporate-site-harness}
}

Contributions

Scripts: datasets/harness-llm/. Issues: GitHub.

Reproduction

python3 datasets/harness-llm/scripts/build_dataset.py
python3 datasets/harness-llm/scripts/validate_dataset.py
python3 datasets/harness-llm/scripts/export_huggingface.py
# optional upload (requires HF token):
python3 datasets/harness-llm/scripts/upload_huggingface.py \
  --repo-id SafetyMP/corporate-site-harness-training-data

How to Use (SFT)

from datasets import load_dataset

ds = load_dataset("SafetyMP/corporate-site-harness-training-data")
# or messages_only:
# ds = load_dataset("SafetyMP/corporate-site-harness-training-data", "messages_only")

def formatting_prompts_func(examples):
    texts = [
        tokenizer.apply_chat_template(
            convo, tokenize=False, add_generation_prompt=False
        )
        for convo in examples["messages"]
    ]
    return {"text": texts}

train_dataset = ds["train"].map(formatting_prompts_func, batched=True)

Pass train_dataset to your SFTTrainer. Keep validation for evaluation only.