Aegis-14B
A governed
dataset-refinery agent, fine-tuned from
NousResearch/Hermes-4-14B (Qwen3-14B base) to power
Aegis — a Conductor-governed autonomous dataset refinery that turns messy data into clean ShareGPT/ChatML training sets, with a human gate on every dollar the agent spends and a signed audit certificate.
Built for the Nous Research × NVIDIA × Stripe hackathon. Trained and quantized locally on a single NVIDIA DGX Spark (GB10).
Note: ~14.8B params (Hermes-4-14B base). The model's system prompt self-identifies as "Aegis-7B" (the project's early working name); the model and repo are Aegis-14B.
What it does
Given a data-refinement task it returns strict JSON for four jobs:
- quality — assess a raw data sample's fitness for fine-tuning
- triage — score a refinement job's complexity / risk
- spend — propose (or, by default, reject in favor of local) a gated external-tool spend
- audit — produce the decision trace for the signed certificate
It is trained to be conservative and local-first: it only proposes paid external work when local processing genuinely can't do the job, and never assumes spend approval.
Usage (OpenAI-compatible)
Serve with vLLM (on GB10, launch with --attention-backend TRITON_ATTN --enforce-eager). Send the Aegis system prompt as the system message; the model replies with JSON only. Use temperature=0 and response_format={"type":"json_object"}.
Training
LoRA (r=16, α=32) over 433 synthetic ShareGPT examples across the four jobs (balanced), 3 epochs, bf16, on one DGX Spark. Held-out eval (48): 100% valid JSON, 91% schema-correct.
License & credits
Derivative of
NousResearch/Hermes-4-14B and inherits its license terms. Credit to
Nous Research (Hermes 4) and the
Qwen3 base. Built with
AInode on
NVIDIA DGX Spark.
