train_chaski.py

#!/usr/bin/env python3
# /// script
# requires-python = ">=3.10"
# dependencies = [
#     "unsloth",
#     "trl>=0.12.0",
#     "peft>=0.7.0",
#     "datasets",
#     "transformers",
#     "huggingface_hub",
#     "trackio",
# ]
# ///
"""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")