doryno/edr-ai-dataset

edr-dataset — EDR-AI training dataset (v3)

Training data for the EDR-AI verdict model (edr-ai, Qwen2.5-3B-Instruct, GGUF, served via LM Studio).

Current release: dataset-v11/ — dataset v3, round 3, used to train the shipped edr-v11 model (97.15% on the frozen 2000-record test split). The legacy v1 corpus behind model v9 stays at the repo root, unchanged.

Model card: https://huggingface.co/doryno/edr-ai-Qwen2.5-3B-Instruct.Q4_K_M

Current release — dataset-v11 (round 3)

Compact v2 schema: each record is {"messages": [{"role": "user", ...}, {"role": "assistant", ...}]} — one Sysmon event per record and a strict JSON verdict. The system prompt is not repeated per record; it is injected at load time from system_prompt.txt and its SHA-256 is pinned in every eval report.

File	Records	Balance	SHA-256
dataset-v11/train.json	9000	2250 per class	b870feff57d11375f7c38539c276d235f9cbec13add47b6734f3572362d05ece
dataset-v11/test.json	2000 (frozen holdout)	500 per class	b477cf000f419b968e846b89ea6cf54be33bf2c30b62ea247118d80d4173c0f6

• Composition: frozen v1 corpus train part (6080 records) + deterministic generator fill to 2250 per class; generator dataset/generator.py v0.3.0, seeds 20260904 (train) / 31137 (holdout) — the same seeds always reproduce byte-identical records.
• Round-3 changes vs v2: S/M boundary relabels (remote mshta.exe HTA staging and process-cradle download-and-execute → MALICIOUS), new hard negatives (local mshta.exe, vendor autostart), a masquerading Sysnative template, and RelatedEvents correlation context on ~55% of generated records (so the production agent can run with EDR_RELATED=1).
• test.json is leak-checked against the train split and excluded from training. Never train on it.

Legacy release (model v9, round 2026-09-01)

Full ChatML dataset — 6400 records, 3 messages each (system = runtime system prompt, user = Sysmon event, assistant = verdict JSON): dataset_chatml_cot.json (6400), dataset_train.json (6080, holdout excluded), test_heldout.json (320, 80 per class — never train on it).

Class balance: BENIGN 2200 / SUSPICIOUS 1800 / MALICIOUS 1600 / MINER 800, 264 scenario families, 8 environment contexts, all inputs unique.

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).

Evaluation history

round	test split	accuracy	note
v8 — 2026-08-27	leaked into training	87.2%	partially memorized, reference only
v9 — 2026-09-01	frozen, 320 records	97.8%	legacy ChatML dataset, shipped as edr-v9
v10 — 2026-09-05	frozen, 2000 records	96.05%	dataset v2 (not released here)
v11 — 2026-09-05	frozen, 2000 records	97.15% (CI 96.35–97.85)	dataset v3 = dataset-v11/, shipped as edr-v11

Live probes (edr-v11 in LM Studio): related-context config (EDR_RELATED=1) 95.95% on 148 probes — the related-context drift of the un-retrained round-2 model (93.24%) is closed by training on dataset-v11/.

Metrics are comparable only within the same protocol: the v8 holdout was leaked (not comparable), v9 used the small 320-record split, rounds v10+ use the full frozen 2000-record split.

Files (repo root, legacy tooling)

File	Purpose
system_prompt.txt	The system prompt — still current, byte-identical across all rounds (SHA-256 pinned in every eval report)
dataset_chatml_cot.json	Legacy v1 corpus — 6400 records, self-validated
dataset_train.json	Legacy v1 training export (6080, holdout excluded)
test_heldout.json	Legacy v1 frozen holdout (320, 80 per class)
edr_agent.py	Runtime agent copy (imported by eval_model.py)
eval_model.py, compare_evals.py	Legacy model-side gate: evaluate a served GGUF against the holdout
run_checks.py	Legacy dataset-side regression gate
make_train_set.py, make_holdout.py, split_1_4.py, inspect_dataset.py	Legacy v1 dataset utilities
thresholds.json	Ship gate: accuracy ≥ 0.95, json_valid ≥ 0.98
eval_reports/	Evaluation runs + round write-ups (v8–v9 here; rounds 2–3 live in the GitHub repo)

The scenario generator and dataset build tooling are published in the edr-ai repo (dataset/generator.py, dataset/build_v3.py, dataset/build_v2.py); deterministic rebuilds from the pinned seeds are byte-identical to the released files. Current training notebooks: notebooks/colab_sft_qwen3b_v3_train.ipynb (A100) and notebooks/colab_sft_qwen3b_v3_eval.ipynb (L4) in the same repo.

Usage

# verify the released files (sha256 must match the pins above)
python - <<'PY'
import hashlib
for f in ("dataset-v11/train.json", "dataset-v11/test.json"):
    print(f, hashlib.sha256(open(f, "rb").read()).hexdigest())
PY

# rebuild / verify from the GitHub repo (deterministic, seeds pinned)
python dataset/build_v3.py check

# legacy v1 gate
python run_checks.py

# evaluate a served build (current: hybrid + error-analysis runners)
python eval_hybrid.py --live 150 --model edr-v11 --related
python eval_errors.py --live 250 --model edr-v11

Round-over-round comparisons: compare_evals.py covers the legacy rounds; rounds 2–3 reports are ingested by ingest_report.py in the GitHub repo.

Export rule (QLoRA → GGUF) — verify every export

The shipped GGUF q8_0 is exported from the merged bf16 composite that passed the frozen-holdout ship gate, and sanity-checked against the same split before release. Do not trust an unverified export path: the v9 round showed that a merge/quantization change can silently destroy MALICIOUS escalation while the output still looks perfectly formatted and confident. Any new build (different base, merge, or quantization) must re-run the full frozen test split before shipping. Full write-up: eval_reports/round_2026-09-01_v9.md.

Contract

• System prompt byte-identical between training data and runtime (system_prompt.txt, SHA-256 b9ce23a9b003d7da1b4f0fafc6b04e78b9003b3e92c01b467dfa5f9104690e64).
• 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 (validate_verdict in edr_agent.py) enforces: JSON shape and 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.

Project

• Runtime agent, rules layer, correlation layer, build tooling, current notebooks and round reports: https://github.com/DoryNo/edr-ai
• Model card: https://huggingface.co/doryno/edr-ai-Qwen2.5-3B-Instruct.Q4_K_M