doryno/edr-ai-dataset

edr-dataset — EDR-AI training dataset (v2)

ChatML training dataset for the EDR-AI verdict model (edr-ai, Qwen2.5-3B-Instruct, GGUF, served via LM Studio). 6400 records — system = runtime system prompt, user = one Sysmon event, assistant = strict JSON verdict (reasoning, classification, risk_score, tactics, techniques, confidence, evidence, recommended_actions).

Class balance: BENIGN 2200 / SUSPICIOUS 1800 / MALICIOUS 1600 / MINER 800, 264 scenario families, 8 environment contexts, all 6400 inputs unique. The set was built to fix the weak spots measured in the previous round: SUSPICIOUS hard negatives across the full 4–8 risk band, BENIGN lookalikes of suspicious shapes, and broader MALICIOUS/MINER diversity.

Evaluation history:

round	holdout	accuracy	note
v8 — 2026-08-27	leaked into training	87.2%	partially memorized, reference only
v9 — 2026-09-01	frozen, leak-checked	97.8%	shipped build, see eval_reports/round_2026-09-01_v9.md

Files (public release)

File	Purpose
dataset_chatml_cot.json	Full dataset — 6400 records, self-validated
dataset_train.json	Training export — the 320 frozen holdout records are excluded (6080 records)
test_heldout.json	Frozen holdout — 320 records, 80 per class; never train on it
system_prompt.txt	The system prompt; must stay byte-identical between training data and runtime
colab_train_edr.ipynb	Reproducible QLoRA training (unsloth, T4) + honest holdout evaluation
run_checks.py	One-shot dataset-side regression gate
make_train_set.py	Refreshes dataset_train.json (holdout excluded) after dataset rebuilds
make_holdout.py	Holdout freezing (do not regenerate between rounds)
split_1_4.py, inspect_dataset.py	Split / inspection utilities
eval_model.py, compare_evals.py	Model-side gate: evaluate a served GGUF against the holdout
thresholds.json	Ship gate: accuracy ≥ 0.95, json_valid ≥ 0.98
eval_reports/	All evaluation runs + round write-ups

Notes:

• The scenario generator is not published — the dataset ships as validated JSON only.
• eval_model.py imports SYS_PROMPT, LM_URL and validate_verdict from edr_agent.py (runtime repo: edr-ai) — copy it next to this folder to run the gate.
• Public release normalization: 101 legacy records carried a real Windows account/machine name captured before the generator went fully synthetic; all identity strings were replaced with a synthetic equivalent. Record counts and classifications are unchanged (see provenance in eval_reports/round_2026-09-01_v9.md).

Usage

python run_checks.py          # dataset-side regression gate
python make_train_set.py      # refresh training export after rebuilds
# train in Colab on dataset_train.json (NOT dataset_chatml_cot.json)
# evaluate a served build:
python eval_model.py --workers 8 --model <lm-studio-model-id>
python compare_evals.py       # round-over-round table

Metrics from rounds before v9 are not comparable: the old holdout was sampled from the same file training used (partial memorization). Since v9 the holdout is frozen and excluded from training.

Export rule (QLoRA → GGUF) — mandatory

Export GGUF only from the training composite: load lora_adapters with unsloth on the bnb-4bit base and call save_pretrained_gguf directly. The adapter compensates the 4-bit quantization error of its training base — merging it onto a full-precision base (plain peft merge_and_unload, or save_pretrained_merged / merged_16bit artifacts) silently destroys MALICIOUS escalation while the output still looks perfectly formatted and confident. Full write-up: eval_reports/round_2026-09-01_v9.md.

Contract

• System prompt byte-identical between training data and runtime.
• Assistant output: one JSON object, fixed key order reasoning, classification, risk_score, tactics, techniques, confidence, evidence, recommended_actions; risk bands BENIGN 1–3, SUSPICIOUS 4–8, MALICIOUS 9–10, MINER 8–10; class-derived action sets.
• Validator enforces: JSON shape/key order, risk bands, ATT&CK technique→tactic mapping, explicit miner evidence, prompt byte-match, no duplicates, no Cyrillic, EventID/EventType consistency, truncation acknowledgement, reasoning template.