Lala L1
Part of the LilaFold research project β small, specialised AI models for serious work.
Lala-L1 is a small causal sequence model for Task A: score each endpoint event as part of a scripted attack step. It was trained by the lilafold lab on owned, real-time Linux lab telemetry (OCSF-normalized events). It is a research checkpoint, not a product, and it does not autonomously block or respond.
- Architecture: 4-layer causal transformer,
d=128, context = last 256 events per host, 3-seed ensemble (~1β3M parameters). - License: CC BY-NC-4.0 (Creative Commons Attribution-NonCommercial 4.0 International).
- Inference:
python src/infer.py --model checkpoints/lala-l1 --input <events.jsonl|campaign dir|-ocsf.tgz> --out scores.csv --fpr 0.001 --calibration host_baseline(train_cvandtrain_cutoffsare also supported). - Full model card:
MODEL_CARD.md. Incident grouping (no retraining):INCIDENTS.md.
Headline results (honest)
| Setting | Result |
|---|---|
| Frozen TEST (12 campaigns, 6 families, eval_freeze_v1) | 0.72 PR-AUC; host-baseline 0.71 recall at 0.44% event FPR |
| 5 unseen campaigns in the frozen-split protocol | 0.80 PR-AUC |
| Small open-weight LLMs (e.g. Qwen-1.5B), same Task A setup | Lala L1 matches or beats them within statistical uncertainty, at a fraction of the compute |
30 fresh real-time validation runs (known families, benign_mix v1, stateful incident grouping) |
30/30 campaigns detected at 1.08 false incidents / host-day |
Frozen TEST and the 30-run check are different evaluations. The 1.08 figure is not a general operating point.
Limitations
- About 14% false alarms on lookalike admin commands (active v1.1 work).
- Five attack families have no real-time frozen TEST data: collection staging, credential access, defense evasion, persistence, privilege escalation.
- On frozen TEST, stateful grouping still leaves 23.44 false incidents / host-day (13/15 campaigns detected).
- Frozen TEST is old-collector telemetry; harder
benign_mix v2mixed-background evaluations are still collecting. - Calibration does not transfer cleanly across collector or benign generations; prefer host-baseline or same-schema real-time calibration.
- No autonomous blocking.
Inference Providers NEW
This model isn't deployed by any Inference Provider. π Ask for provider support