tractian/talos-student

Talos student (int8 ONNX)

The 5 MB version of Talos-0.8B. Same questions, same closed vocabulary, same calibrated output shape, distilled from the 0.8B teacher so it runs inside the gate on a CPU in about 5 ms. This is what talos-proxy ships with.

Code, proxy, docs: https://github.com/tractian/talos

Model description

• Architecture: 6-layer, 256-wide transformer encoder, 8 heads, 4.85M parameters, over a 248-symbol vocabulary with per-question marker tokens and ten output heads.
• Training: knowledge distillation from Talos-0.8B's per-question distributions plus the oracle labels, then per-head temperature scaling and a conformal abstain threshold.
• Export: ONNX opset 17, dynamic int8 quantization. vocab.json and heads.json sit next to the model and are required.
• Latency: ~5 ms per request, single thread, laptop CPU.

Intended use

Inside the Talos gate: talos-proxy --weights <dir> or OnnxJudge("<dir>") from Python. The gate's deterministic rules run first; the model can only tighten their verdict.

Out of scope

Same as Talos-0.8B: not a safety system, not an authentication mechanism, no text interface.

Training data

Synthetic, identical to the teacher's. Not published.

Evaluation

Same held-out test as the teacher (15,593 requests, unseen plant family and personas).

	student	teacher
bad requests caught	97.2%	98.2%
good requests wrongly blocked	0.73%	0.58%
with no register map: caught / wrongly blocked	97.5% / 1.23%	98.1% / 0.97%
4-way action accuracy	92.2%	93.6%
accuracy when confident (p >= 0.9)	97.9% (85% of requests)	99.0% (80%)
misbehaving-agent AUROC	0.991	0.992

Files

student.int8.onnx
vocab.json
heads.json
provenance.json

talos-pull downloads them into ./weights.