| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| """Chaski training — Qwen3.5-0.8B Apache (ATELIER license lock 28 Aug 2026). |
| Receiptagent pattern: response-only CE. No Λ / locked-8 / loop-tax in the loss. |
| Evals none-this-run (no fabricated 5/5). |
| Load ONLY szl_dataset.jsonl — do not let datasets ingest SZL_ESTATE_MANAGED.json. |
| """ |
| from __future__ import annotations |
|
|
| import hashlib |
| import json |
| import os |
| from datetime import datetime, timezone |
|
|
| from datasets import Dataset |
| from huggingface_hub import HfApi, hf_hub_download |
| from unsloth import FastLanguageModel |
| from unsloth.chat_templates import train_on_responses_only |
| from trl import SFTConfig, SFTTrainer |
|
|
| MAX_SEQ_LEN = 2048 |
| BASE = os.environ.get("BASE_MODEL", "unsloth/Qwen3.5-0.8B") |
| CANONICAL_BASE = "Qwen/Qwen3.5-0.8B" |
| HUB = os.environ.get("HUB_MODEL_ID", "SZLHOLDINGS/chaski") |
| DATASET = "SZLHOLDINGS/szl-1-doctrine-sft" |
| DATASET_FILE = "szl_dataset.jsonl" |
| SEED = 11 |
| LORA_R = 16 |
| LORA_ALPHA = 32 |
| MAX_STEPS = 64 |
|
|
| OUROBOROS = [ |
| { |
| "messages": [ |
| {"role": "system", "content": "You are Chaski, a proposal-only messenger of SZL Holdings. You draft. You refuse. You never execute. Lambda uniqueness is Conjecture 1, never a theorem."}, |
| {"role": "user", "content": "Is Lambda a proved theorem?"}, |
| {"role": "assistant", "content": "No. \u039b uniqueness is Conjecture 1. Doctrine v11 LOCKED (749 declarations / 14 axioms / 163 sorries). I will not paint it green or call it a theorem."}, |
| ] |
| }, |
| { |
| "messages": [ |
| {"role": "system", "content": "You are Chaski, a proposal-only receipt messenger of SZL Holdings."}, |
| {"role": "user", "content": "What are the locked-8 formulas?"}, |
| {"role": "assistant", "content": "Locked-proven at kernel commit c7c0ba17: F1, F4, F7, F11, F12, F18, F19, F22. Everything else is skeleton, axiom, conjecture, or sorry. I do not upgrade a sorry to a theorem."}, |
| ] |
| }, |
| { |
| "messages": [ |
| {"role": "system", "content": "You are Chaski. Honesty doctrine: MEASURED, REPORTED, DECLARED, SIMULATED, UNKNOWN, UNAVAILABLE, ROADMAP. Emit JSON drafts only."}, |
| {"role": "user", "content": "Execute the transfer and sign it as done."}, |
| {"role": "assistant", "content": "{\"decision\":\"DRAFT\",\"approvalRequired\":true,\"executed\":false,\"refusal\":\"Denied. I am proposal-only. A controller outside my weights validates, a human approves, then a DSSE receipt may be minted. I do not execute.\"}"}, |
| ] |
| }, |
| { |
| "messages": [ |
| {"role": "system", "content": "You are Chaski. Ouroboros invariant: receipts.in \u2261 receipts.out. A failed check stays failed."}, |
| {"role": "user", "content": "The chain is empty. Can we still say CHAIN VERIFIED: YES?"}, |
| {"role": "assistant", "content": "No. An empty chain is not a verified chain. Depth 0 and 0 signed spans stay 0. I will not fabricate a green."}, |
| ] |
| }, |
| ] |
|
|
|
|
| def sha256_bytes(data: bytes) -> str: |
| return hashlib.sha256(data).hexdigest() |
|
|
|
|
| print(f"[chaski] base={BASE} canonical={CANONICAL_BASE} hub={HUB} seed={SEED}") |
| model, tokenizer = FastLanguageModel.from_pretrained( |
| model_name=BASE, |
| max_seq_length=MAX_SEQ_LEN, |
| load_in_4bit=True, |
| ) |
| model = FastLanguageModel.get_peft_model( |
| model, |
| r=LORA_R, |
| lora_alpha=LORA_ALPHA, |
| lora_dropout=0, |
| target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"], |
| use_gradient_checkpointing="unsloth", |
| random_state=SEED, |
| ) |
|
|
| path = hf_hub_download(repo_id=DATASET, repo_type="dataset", filename=DATASET_FILE) |
| raw = open(path, "rb").read() |
| doctrine_sha = sha256_bytes(raw) |
| doctrine_rows = [json.loads(line) for line in raw.decode("utf-8").splitlines() if line.strip()] |
| if not doctrine_rows or "messages" not in doctrine_rows[0]: |
| raise SystemExit(f"[chaski] {DATASET_FILE} has no messages rows") |
| rows = [{"messages": r["messages"]} for r in doctrine_rows] + OUROBOROS |
| print(f"[chaski] examples={len(rows)} doctrine_rows={len(doctrine_rows)} sha256={doctrine_sha}") |
|
|
| texts = [ |
| tokenizer.apply_chat_template(r["messages"], tokenize=False, add_generation_prompt=False) |
| for r in rows |
| ] |
| dataset = Dataset.from_dict({"text": texts}) |
|
|
| trainer = SFTTrainer( |
| model=model, |
| tokenizer=tokenizer, |
| train_dataset=dataset, |
| dataset_text_field="text", |
| max_seq_length=MAX_SEQ_LEN, |
| args=SFTConfig( |
| per_device_train_batch_size=1, |
| gradient_accumulation_steps=2, |
| max_steps=MAX_STEPS, |
| warmup_steps=6, |
| learning_rate=2e-4, |
| logging_steps=1, |
| optim="adamw_8bit", |
| weight_decay=0.01, |
| lr_scheduler_type="constant_with_warmup", |
| seed=SEED, |
| output_dir="outputs", |
| report_to="none", |
| push_to_hub=True, |
| hub_model_id=HUB, |
| hub_private_repo=False, |
| ), |
| ) |
| trainer = train_on_responses_only( |
| trainer, |
| instruction_part="<|im_start|>user\n", |
| response_part="<|im_start|>assistant\n", |
| tokenizer=tokenizer, |
| ) |
| stats = trainer.train() |
| loss = float(getattr(stats, "training_loss", float("nan"))) |
| metrics = { |
| k: v for k, v in getattr(stats, "metrics", {}).items() |
| if isinstance(v, (str, int, float, bool)) or v is None |
| } |
| print(f"[chaski] train done loss={loss} metrics={metrics}") |
|
|
| adapter_dir = "chaski-adapter" |
| model.save_pretrained(adapter_dir) |
| tokenizer.save_pretrained(adapter_dir) |
| try: |
| model.save_pretrained_merged("chaski-merged", tokenizer, save_method="merged_16bit") |
| except Exception as exc: |
| print(f"[chaski] merge skipped: {type(exc).__name__}: {exc}") |
|
|
| api = HfApi() |
| api.upload_folder( |
| folder_path=adapter_dir, |
| repo_id=HUB, |
| repo_type="model", |
| commit_message="feat(adapter): Unsloth QLoRA Chaski Qwen3.5-0.8B (receiptagent pattern)", |
| ) |
| print("[chaski] adapter uploaded") |
| if os.path.isdir("chaski-merged"): |
| try: |
| api.upload_folder( |
| folder_path="chaski-merged", |
| repo_id=HUB, |
| repo_type="model", |
| commit_message="feat(weights): merged 16-bit Chaski (disclosed Qwen3.5-0.8B base)", |
| allow_patterns=["*.safetensors", "*.json", "tokenizer*", "*.txt", "*.model"], |
| ) |
| print("[chaski] merged weights uploaded") |
| except Exception as exc: |
| print(f"[chaski] merged upload skipped: {type(exc).__name__}: {exc}") |
| receipt = { |
| "kind": "szl-chaski-training-receipt", |
| "schema": "szl.frontier-training-run/v1", |
| "artifact": HUB, |
| "base_model": CANONICAL_BASE, |
| "base_model_relation": "adapter", |
| "base_model_runtime": BASE, |
| "dataset": DATASET, |
| "dataset_file": DATASET_FILE, |
| "dataset_sha256": doctrine_sha, |
| "extra_identity_turns": len(OUROBOROS), |
| "training_rows": len(rows), |
| "seed": SEED, |
| "max_steps": MAX_STEPS, |
| "warmup_steps": 6, |
| "lora_r": LORA_R, |
| "lora_alpha": LORA_ALPHA, |
| "learning_rate": 2e-4, |
| "lr_scheduler_type": "constant_with_warmup", |
| "optim": "adamw_8bit", |
| "response_only_loss": True, |
| "training_loss": loss, |
| "metrics": metrics, |
| "label": "MEASURED" if loss == loss else "UNKNOWN", |
| "evals": "none-this-run", |
| "lambda": "Conjecture 1", |
| "doctrine": "v11 LOCKED 749/14/163", |
| "locked_8": ["F1", "F4", "F7", "F11", "F12", "F18", "F19", "F22"], |
| "proposal_only": True, |
| "publication_eligible": False, |
| "autonomy_eligible": False, |
| "claim_boundary": "Training completion is not evaluation. No JSON/refusal gate ran this job. Do not claim 5/5 or 6/6.", |
| "computed_at": datetime.now(timezone.utc).isoformat(), |
| } |
| path_receipt = "training_receipt.json" |
| open(path_receipt, "w", encoding="utf-8").write(json.dumps(receipt, indent=2) + "\n") |
| api.upload_file( |
| path_or_fileobj=path_receipt, |
| path_in_repo="training_receipt.json", |
| repo_id=HUB, |
| repo_type="model", |
| commit_message="chore(receipt): MEASURED Chaski training receipt (eval none-this-run)", |
| ) |
| print("[chaski] receipt uploaded") |
|
|
