szl-training-scripts

The forge. Hub copy of the Unsloth-receipted training scripts. Tags were empty — this atelier is the card.

Family. estate · Evidence. HUB · Weights. none (UNAVAILABLE on this Hub ID: scripts and receipts only)

Hub: SZLHOLDINGS/szl-training-scripts

Canonical source (GitHub): the atelier kit lives in szl-holdings/a11oy under .github/atelier-kit/ (receipted_unsloth.py, unsloth-recipes.json, cards/szl-training-scripts.md, kit-README.md). The kit README names szl-holdings/szl-khipu, szl-holdings/szl-forge and szl-holdings/szl-nemo as its upstream sources. GitHub is canonical; this Hub repository is the mirror.

Hub tree (OBSERVED 2026-09-25)

Listed through the Hub file API; the revision SHA was not captured by this pass. No weight file of any kind is present.

Path	Bytes	What it is (from the file's own header)
receipted_unsloth.py	12,165	"Receipted Unsloth — take the speed, bind the knobs, refuse the folklore." Unsloth QLoRA trainer with one PROFILES entry per organ (khipu, khipu-r2, receiptagent, willay, chaski, chaski-5050, chaski-r2), house seed 20260721, --receipt-only (mint the recipe receipt, no GPU) and --list. Same file name and byte size as the a11oy atelier-kit copy; the first ~3 KB were compared and match, the full-file digest was not compared.
train_receipted_unsloth.py	12,431	"Receipted Unsloth unique-cut GPU trainer. Runs on Hugging Face Jobs (A10G)." uv run script header; --profile willay / --profile chaski.
train_willay.py	6,799	WILLAY signed-refusal SFT LoRA script (base Qwen/Qwen2.5-0.5B-Instruct, target SZLHOLDINGS/WILLAY).
train_cohort.py	17,351	"CPU-train WILLAY + Chaski silhouettes. Measure kernels. No 1.5B retrain." NumPy silhouettes, seed 20260721, --kernels-only.
szl_khipu_eval.py	10,412	KHIPU-R2 held-out eval port (eval.jsonl 5 navigate + adversarial.jsonl 6; fp16 and optional 4-bit NF4 comparison).
unsloth-recipes.json	16,247	recipe table (same size as the atelier-kit copy).
receipts/chaski-unsloth.json, receipts/chaski-5050-unsloth.json, receipts/chaski-r2-unsloth.json, receipts/willay-unsloth.json	1,592 / 1,638 / 1,587 / 1,641	szl.training_receipt.v2 receipts minted by the profile trainer (unsigned).
training_receipt.unsloth.json	1,638	profile chaski-5050 receipt (path_in_repo: adapter-unsloth, train loss 0.5588 MEASURED, evals: none-this-run, energy UNAVAILABLE, proven_trust: false).
provenance.json / status.json / LICENSE	1,755 / 1,394 / 11,357	license-candidate records and standalone Apache-2.0 text. status.json keeps lineage, consent, privacy, training-suitability and deployment fields BLOCKED and production_ready: false.
bom/model-bom.cdx.json	5,324	CycloneDX BOM
.gitattributes	1,519	platform-managed

The cut

Training code is usually a gist. We give it a model id so the estate graph stays one piece.

A Hub id you can pin in a receipt: 'trained by this repo at this SHA'.

Silhouette → leave → SZL

Leader	Take, then tweak
Anthropic	No public train scripts for Claude.
NVIDIA	NeMo recipes.
Unsloth	The scripts wrap Unsloth v2 profiles. Cut is unique knobs per organ, bound in the receipt.

No equivalent public combination was found among the leaders surveyed above (a snapshot, not an ecosystem-wide novelty claim).

Intended use

Canonical train entrypoint.

Limitations

• Not a checkpoint. Wire to GitHub kit in this atelier.
• The receipts here are unsigned recipe/run envelopes; a receipt on this Hub ID does not by itself bind any adapter bytes on another repository.

Honesty

Claim	Label
This card's numbers	HUB
Energy / joules	UNAVAILABLE unless a signed meter says MEASURED
Λ uniqueness	Conjecture 1 OPEN — not a theorem
GGUF as the signed object	FALSE

Doctrine v11 LOCKED · 749 declarations · 14 axioms · 163 sorries · locked-proven 8.

Apache-2.0. Copyright 2026 SZL Holdings · Stephen P. Lutar Jr. · ORCID 0009-0001-0110-4173.

GitHub-aligned Python

An earlier version of this card embedded a code block that stopped mid-statement (if trainer.state.log_hi) and that did not match the receipted_unsloth.py actually stored on this Hub ID (it showed an older 2,932-byte single-profile variant with --base/--r/--seed flags, which still exists as atelier/receipted_unsloth.py in szl-khipu and szl-forge). That block is removed. What follows is quoted verbatim from the 12,165-byte receipted_unsloth.py on this repository (2026-09-25 read).

Usage lines from the module docstring:

  python receipted_unsloth.py --profile willay --receipt-only
  python receipted_unsloth.py --profile chaski --data doctrine.jsonl --out out/chaski

The receipt-minting tail of main() — the point of the file: every knob lands in training_receipt.json, loss is read from the trainer log, and the envelope is meant to be signed before any merge:

    trainer = SFTTrainer(model=model, tokenizer=tokenizer, train_dataset=ds, args=SFTConfig(**sft_kw))
    t0 = time.time()
    trainer.train()
    Path(args.out).mkdir(parents=True, exist_ok=True)
    model.save_pretrained(args.out)
    loss = None
    if trainer.state.log_history:
        last = trainer.state.log_history[-1]
        loss = last.get("train_loss") or last.get("loss")
    extra.update({"final_loss": loss, "seconds": round(time.time() - t0, 3), "honesty": "REPORTED"})
    rec = mint_receipt(args.profile, cfg, extra)
    Path(args.out, "training_receipt.json").write_text(json.dumps(rec, indent=2) + "\n")
    print(json.dumps(rec, indent=2))
    return 0

mint_receipt refuses proven_trust: true and refuses any non-null energy_j ("refusing to fabricate joules"); the receipt records "gguf": "derived — never the signed object". Read the full file on this repository or in the canonical kit linked above; do not treat this excerpt as the whole program.