osava-smollm (SENTRI-Classifier v7b)

A QLoRA fine-tune of SmolLM3-3B, quantised to Q4_K_M GGUF, that rates one Windows Sysmon or Security event as none / low / medium / high. It is the local triage model of OSAVA's Behaviour tab: it decides which behaviours a person should look at first. It triages; it does not prove. A High means "look here first", not "this is an attack".

File	SHA-256	Bytes
osava-smollm-q4.gguf	dc24b6b923ed4dc7b475f41f31a96a90a1cfa242241e35ea37de09ce10c11f52	1,915,305,376

OSAVA pins this file by checksum in its contract.json and refuses any other.

How it is used

One event, rendered as Key: value lines by a fixed contract (field order, 512-character command-line limit, signature backfill), goes into a fixed prompt (strict variant, prompt SHA-256 8906351f…) that ends with the prefill {"severity": ". Only the first generated token is read: its top-k probabilities over the four severity words give the verdict and how close the call was. Everything else a person sees (indicators, threat type, reasoning) is computed from the event by fixed rules, not generated. The prompt, conversion rules and canary events are in OSAVA's contract.json and in osava-slm, which built this model.

llama-server -m osava-smollm-q4.gguf
POST /completion  {"prompt": <template with the event>, "n_predict": 1, "n_probs": 10, "temperature": 0}

Training

• Base: SmolLM3-3B. QLoRA (NF4, r=16, alpha=32, all linear projections), 4 epochs, learning rate 2e-4, effective batch 16, loss on the completion only with the severity token weighted 20x. Final checkpoint (580 steps), not the best of several.
• Data (2,522 rows):
  • Events from EVTX-ATTACK-SAMPLES (GPL-3.0), labelled by behaviour.
  • Synthetic events from generators in osava-slm: benign hard negatives, suspicious grey-zone activity, logons, and (new in v7b) normal workstation activity, each family paired with malicious contrasts: installers and updaters, per-user Windows services, app packages, per-user apps on HTTPS, developer tools.
  • Rebalanced to 42.5% malicious, 16.7% suspicious, 40.8% benign by repeating existing rows.
  • No telemetry from any real user's machine is in the training data.

Evaluation

Measured the way OSAVA runs it (this GGUF through llama-server, first-token readout):

	v3 (previous)	v7b
Held-out events, accuracy (138)	124	122
Held-out malicious events called High (56)	55	53 (the other 2 Medium)
Held-out suspicious events caught	53%	67%
A working PC's benign behaviours called High, applications never seen in training (483)	26.1%	9.3%
Same, High or Medium	33.1%	20.1%

Caveats: the held-out set is small, and v7b's data change was chosen after seeing an earlier candidate's misses on it, so it is not fully blind for this model. The benign benchmark comes from one developer's machine. Both v3 and v7b call one held-out malicious event benign.

Limits

• One event at a time, no memory: it cannot see a sequence (download, then execute) that only looks bad together.
• The grey zone (Medium) is its weakest call; read None and Low as "nothing stood out", not "safe".
• It was trained on Sysmon event IDs 1, 3, 7, 11, 12, 13 and Security 4624/4625. Anything else is out of its domain.
• Attackers who make their activity look like ordinary software (a plausibly named service, a signed tool doing its normal job) can be rated low.

Licence

The weights are released under Apache-2.0, the base model's licence. Part of the training data comes from EVTX-ATTACK-SAMPLES, which is GPL-3.0; whether a dataset's licence carries over to weights trained on it is not settled, and you should take that into account for your use.