Add safe log-to-training demo pipeline

[aayushprsingh]

Summary

Adds a small, dependency-free demo pipeline for the DMP 2026 logs-to-training project.

The demo shows the expected end-to-end flow on toy agent logs:

• ingest JSONL agent/Q&A sessions
• redact basic PII and secrets
• preserve tool calls + tool observations as trajectory data
• emit SFT/LoRA-style JSONL
• emit DPO-style JSONL
• tag trajectory complexity
• write a validation report with PII counts and schema errors

Files

• scripts/log_to_training_data.js — demo pipeline
• samples/agent_logs.jsonl — toy Q&A + agentic logs
• schemas/training_records.schema.json — minimal schema contract
• out/*.jsonl and out/validation_report.json — generated example artifacts
• DEMO.md — run instructions and design notes

Verification

Ran locally:

node scripts/log_to_training_data.js samples/agent_logs.jsonl out

Output:

Processed 2 sessions
PII redactions: {"email":1,"phone":1,"name":1,"api_key":1}
Wrote out/sft.jsonl, out/dpo.jsonl, out/validation_report.json

out/validation_report.json shows zero validation errors.

Note

This is intentionally small so mentors can review direction quickly. If this aligns with the project direction, next step can be making schema validation stricter and separating redaction, segmentation, and export into modules.