# SparkLang live train — three CPU methods (not LoRA)
#
# Captured 2026-09-02 against tools/spark-train-ref on :8090
# Proof: real PyTorch state_dicts / JSON packs — not marker stubs.

## Methods listed by trainer
{"ok":true,"service":"spark-train-ref","methods":["spark_distill_cpu","spark_pref_pack","spark_playbook_fit","spark_faq_index"]}


## 1) spark_distill_cpu — ./spark --live examples/model_train.spark
[spark] live ask via AI_GATEWAY_URL (spark-ask-http)
[model] code
[model] {"job_id":"job-dry-001","status":"accepted","artifacts":{"checkpoint":"out/train/job-dry-001/checkpoint.json","marker":"out/train/job-dry-001/ARTIFACT","adapter":"out/train/job-dry-001/weights.pt","weights":"out/train/job-dry-001/weights.pt"},"out":"out/train/job-dry-001","base":"fixture-base","dataset":"examples/fixtures/train/dataset.jsonl","backend":"http","method":"spark_distill_cpu","note":"spark_distill_cpu \u2014 CPU train (loss=1e-06; not LoRA)","final_loss":1e-06,"methods":["spark_distill_cpu","spark_pref_pack","spark_playbook_fit","spark_faq_index"]}
[model] {"job_id":"job-dry-001","state":"succeeded","artifacts":{"checkpoint":"out/train/job-dry-001/checkpoint.json","marker":"out/train/job-dry-001/ARTIFACT","adapter":"out/train/job-dry-001/weights.pt","weights":"out/train/job-dry-001/weights.pt"},"out":"out/train/job-dry-001","base":"fixture-base","dataset":"examples/fixtures/train/dataset.jsonl","backend":"http","method":"spark_distill_cpu","note":"spark_distill_cpu \u2014 CPU train (loss=1e-06; not LoRA)","final_loss":1e-06}
[spark] ok

### Artifact sizes
total 24
drwxrwxr-x 2 mike mike 4096 Sep  2 07:12 .
drwxr-xr-x 8 mike mike 4096 Sep  2 07:12 ..
-rw-rw-r-- 1 mike mike  119 Sep  2 07:12 ARTIFACT
-rw-rw-r-- 1 mike mike  409 Sep  2 07:12 checkpoint.json
-rw-rw-r-- 1 mike mike 5601 Sep  2 07:12 weights.pt

## 2) spark_pref_pack — companion --method
{"job_id":"job-pref-001","status":"accepted","artifacts":{"checkpoint":"out/train/job-pref-001/checkpoint.json","marker":"out/train/job-pref-001/ARTIFACT","adapter":"out/train/job-pref-001/ranker.pt","weights":"out/train/job-pref-001/ranker.pt","pref_pack":"out/train/job-pref-001/pref_pack.json","ranker":"out/train/job-pref-001/ranker.pt"},"out":"out/train/job-pref-001","base":"spark_pref_pack","dataset":"examples/fixtures/train/dataset.jsonl","backend":"http","method":"spark_pref_pack","note":"spark_pref_pack \u2014 CPU train (loss=0.492453; not LoRA)","final_loss":0.492453,"methods":["spark_distill_cpu","spark_pref_pack","spark_playbook_fit","spark_faq_index"]}
{"job_id":"job-pref-001","state":"succeeded","artifacts":{"checkpoint":"out/train/job-pref-001/checkpoint.json","marker":"out/train/job-pref-001/ARTIFACT","adapter":"out/train/job-pref-001/ranker.pt","weights":"out/train/job-pref-001/ranker.pt","pref_pack":"out/train/job-pref-001/pref_pack.json","ranker":"out/train/job-pref-001/ranker.pt"},"out":"out/train/job-pref-001","base":"spark_pref_pack","dataset":"examples/fixtures/train/dataset.jsonl","backend":"http","method":"spark_pref_pack","note":"spark_pref_pack \u2014 CPU train (loss=0.492453; not LoRA)","final_loss":0.492453}

total 28
drwxrwxr-x 2 mike mike 4096 Sep  2 07:12 .
drwxr-xr-x 8 mike mike 4096 Sep  2 07:12 ..
-rw-rw-r-- 1 mike mike  138 Sep  2 07:12 ARTIFACT
-rw-rw-r-- 1 mike mike  447 Sep  2 07:12 checkpoint.json
-rw-rw-r-- 1 mike mike 1207 Sep  2 07:12 pref_pack.json
-rw-rw-r-- 1 mike mike 5592 Sep  2 07:12 ranker.pt

## 3) spark_playbook_fit — companion --method
{"job_id":"job-play-001","status":"accepted","artifacts":{"checkpoint":"out/train/job-play-001/checkpoint.json","marker":"out/train/job-play-001/ARTIFACT","adapter":"out/train/job-play-001/router.pt","weights":"out/train/job-play-001/router.pt","playbooks":"out/train/job-play-001/playbooks.json","router":"out/train/job-play-001/router.pt"},"out":"out/train/job-play-001","base":"spark_playbook_fit","dataset":"examples/fixtures/train/dataset.jsonl","backend":"http","method":"spark_playbook_fit","note":"spark_playbook_fit \u2014 CPU train (loss=3e-06; not LoRA)","final_loss":3e-06,"methods":["spark_distill_cpu","spark_pref_pack","spark_playbook_fit","spark_faq_index"]}
{"job_id":"job-play-001","state":"succeeded","artifacts":{"checkpoint":"out/train/job-play-001/checkpoint.json","marker":"out/train/job-play-001/ARTIFACT","adapter":"out/train/job-play-001/router.pt","weights":"out/train/job-play-001/router.pt","playbooks":"out/train/job-play-001/playbooks.json","router":"out/train/job-play-001/router.pt"},"out":"out/train/job-play-001","base":"spark_playbook_fit","dataset":"examples/fixtures/train/dataset.jsonl","backend":"http","method":"spark_playbook_fit","note":"spark_playbook_fit \u2014 CPU train (loss=3e-06; not LoRA)","final_loss":3e-06}

total 28
drwxrwxr-x 2 mike mike 4096 Sep  2 07:12 .
drwxr-xr-x 8 mike mike 4096 Sep  2 07:12 ..
-rw-rw-r-- 1 mike mike  138 Sep  2 07:12 ARTIFACT
-rw-rw-r-- 1 mike mike  468 Sep  2 07:12 checkpoint.json
-rw-rw-r-- 1 mike mike  918 Sep  2 07:12 playbooks.json
-rw-rw-r-- 1 mike mike 5528 Sep  2 07:12 router.pt

## checkpoint method stamps
out/train/job-dry-001/checkpoint.json {
  "method": "spark_distill_cpu",
  "job_id": "job-dry-001",
  "base": "fixture-base",
  "dataset": "examples/fixtures/train/dataset.jsonl",
  "pairs": 12,
  "vocab_size": 19,
  "n_replies": 5,
  "steps": 200,
  "final_loss": 1e-06,
  "device": "cpu",
  "elapsed_s": 0.194,
  "weights": "out/train/job-dry-001/weights.pt",
  "note": "Spark custom path: CPU distill student (not LoRA, not HF PEFT, no GPU)"
}
out/train/job-pref-001/checkpoint.json {
  "method": "spark_pref_pack",
  "job_id": "job-pref-001",
  "base": "spark_pref_pack",
  "dataset": "examples/fixtures/train/dataset.jsonl",
  "pairs": 12,
  "vocab_size": 23,
  "steps": 200,
  "final_loss": 0.492453,
  "device": "cpu",
  "elapsed_s": 0.107,
  "pref_pack": "out/train/job-pref-001/pref_pack.json",
  "ranker": "out/train/job-pref-001/ranker.pt",
  "note": "Spark method: preference pack + CPU ranker (not LoRA, not HF PEFT)"
}
out/train/job-play-001/checkpoint.json {
  "method": "spark_playbook_fit",
  "job_id": "job-play-001",
  "base": "spark_playbook_fit",
  "dataset": "examples/fixtures/train/dataset.jsonl",
  "pairs": 12,
  "n_playbooks": 5,
  "vocab_size": 19,
  "steps": 200,
  "final_loss": 3e-06,
  "device": "cpu",
  "elapsed_s": 0.063,
  "playbooks": "out/train/job-play-001/playbooks.json",
  "router": "out/train/job-play-001/router.pt",
  "note": "Spark method: playbook fit router on CPU (not LoRA, not HF PEFT)"
}

## GET /v1/
{"ok":true,"service":"spark-train-ref","methods":["spark_distill_cpu","spark_pref_pack","spark_playbook_fit","spark_faq_index"]}


## 4) spark_faq_index — companion --method
{"job_id":"job-faq-001","status":"accepted","artifacts":{"checkpoint":"out/train/job-faq-001/checkpoint.json","marker":"out/train/job-faq-001/ARTIFACT","adapter":"out/train/job-faq-001/encoder.pt","weights":"out/train/job-faq-001/encoder.pt","faq_index":"out/train/job-faq-001/faq_index.json","encoder":"out/train/job-faq-001/encoder.pt"},"out":"out/train/job-faq-001","base":"spark_faq_index","dataset":"examples/fixtures/train/dataset.jsonl","backend":"http","method":"spark_faq_index","note":"spark_faq_index \u2014 CPU train (loss=0.758859; not LoRA)","final_loss":0.758859,"methods":["spark_distill_cpu","spark_pref_pack","spark_playbook_fit","spark_faq_index"]}

{"job_id":"job-faq-001","state":"succeeded","artifacts":{"checkpoint":"out/train/job-faq-001/checkpoint.json","marker":"out/train/job-faq-001/ARTIFACT","adapter":"out/train/job-faq-001/encoder.pt","weights":"out/train/job-faq-001/encoder.pt","faq_index":"out/train/job-faq-001/faq_index.json","encoder":"out/train/job-faq-001/encoder.pt"},"out":"out/train/job-faq-001","base":"spark_faq_index","dataset":"examples/fixtures/train/dataset.jsonl","backend":"http","method":"spark_faq_index","note":"spark_faq_index \u2014 CPU train (loss=0.758859; not LoRA)","final_loss":0.758859}

## ARTIFACT
spark_faq_index job-faq-001
primary=out/train/job-faq-001/encoder.pt
final_loss=0.758859
faq_index=out/train/job-faq-001/faq_index.json

## checkpoint.json (method)
spark_faq_index
## encoder.pt method + keys
spark_faq_index ['emb.weight', 'proj.bias', 'proj.weight']
## faq_index.json n_faqs
5