Chat response distillation and SFT

This tutorial freezes one canonical conversation, produces explicitly fake teacher output, trains the CPU fixture, resumes it, evaluates it, and verifies its export. It makes no network request, model download, GPU call, or upload.

1. Inspect and freeze the input

examples/chat-offline/records.jsonl contains one versioned conversation with provenance and a stable record ID. Freeze it before generation so later stages can bind to an immutable hash.

npm run build

node dist/cli/index.js dataset freeze \
  examples/chat-offline/records.jsonl \
  --out tmp/chat-offline/frozen \
  --force \
  --json

Expect recordCount: 1. The frozen manifest is the input authority; changing a record creates a different hash rather than silently reusing prior work.

2. Initialize, plan, and generate offline

node dist/cli/index.js distill init \
  --root tmp/chat-offline/distill \
  --config examples/chat-offline/distillation.json \
  --input examples/chat-offline/records.jsonl \
  --force \
  --json

node dist/cli/index.js distill plan \
  --root tmp/chat-offline/distill \
  --json

node dist/cli/index.js distill responses \
  --root tmp/chat-offline/distill \
  --offline-fake \
  --json

The fake candidate is labeled custom/offline-fake; it is test data, not evidence about a real teacher. A provider-backed plan instead requires an explicit provider/model, credential environment-variable name, --allow-network, and separate positive generation and judging budgets. Idempotency and the resume ledger prevent a successful paid sample from being purchased twice.

3. Freeze distilled output

node dist/cli/index.js distill freeze \
  --root tmp/chat-offline/distill \
  --out tmp/chat-offline/distilled \
  --force \
  --json

Inspect rejected and retained candidates before freezing. Policy decisions annotate candidates; they do not erase lineage. Keep validation/test groups locked and never send held-out answers to a teacher.

4. Train, resume, evaluate, and export

node dist/cli/index.js training run \
  --spec examples/chat-offline/training.json \
  --python python3 --python-root python --json

node dist/cli/index.js training resume \
  --spec examples/chat-offline/training.json \
  --checkpoint ../tmp/chat-offline/train/checkpoint-1.json \
  --python python3 --python-root python --json

node dist/cli/index.js training evaluate \
  --spec examples/chat-offline/training.json \
  --python python3 --python-root python --json

node dist/cli/index.js training export \
  --spec examples/chat-offline/training.json \
  --python python3 --python-root python --json

Each command returns exitCode: 0. Verify every path and SHA-256 in tmp/chat-offline/train/artifact-manifest.json. A complete checkpoint permits full resume; weights-only output is a labeled warm start.

Recovery and production transition

On interruption, rerun status and resume from the newest atomic, compatible checkpoint. Never adopt a checkpoint with a different data hash, recipe identity, image, or spec. The Qwen 9B recipe and all production chat recipes are experimental; execution remains gated until pinned license, GPU smoke, clean reload, and evaluation gates pass.

Next: choose a distillation target or interpret evaluation and artifacts.