BSLBSL/ai-lab-m2-agent

ai-lab-m2-agent — Month 2 (generic-code, free tier)

Generic prompt → SmolAgents CodeAgent + PythonExecutorTool + DatasetTool → {code, stdout, latency}. Free tier only: openai/gpt-oss-20b via Groq (3 models Inference Available) via hf-inference (toggle ON) or local SmolLM2-360M-Instruct (~700MB CPU).

Tools

• —python_interpreter (PythonExecutorTool, 10s timeout, no network) — fibonacci, csv filter, plot sin, prime 97, etc.
• —dataset_search (DatasetTool, reads BSLBSL/month1-spam-sample 50 rows) — keyword/limit → filtered spam SMS for dataset-aware prompts (e.g. count FREE).

Demos

• —VPS (live): http://51.79.71.160.sslip.io/agent/ and POST /agent/run via Caddy handle_path /agent* → host.docker.internal:8001 (host FastAPI, like /chat)
• —Static fallback: index.html calls POST /agent/run with JS; works with or without HF token.

API

curl -s http://51.79.71.160:8001/health
curl -s -X POST http://51.79.71.160:8001/agent/run -H 'Content-Type: application/json' -d '{"prompt":"fibonacci 20"}'
curl -s -X POST http://51.79.71.160:8001/agent/run -H 'Content-Type: application/json' -d '{"prompt":"count spam SMS containing FREE in BSLBSL/month1-spam-sample"}'

Run locally

pip install -r requirements.txt
HF_TOKEN=hf_xxx REDIS_URL=redis://localhost:6379/0 uvicorn app:app --port 8001

Metrics

Target ≥85% (17/20) generic-code stdout==expected, p50 <4s local 360M / <1s Groq via HF, tokens avg <400, tool calls avg 1.5–2.5. Cost $0. See METRICS.md.