feat(train): Qwen3.5-0.8B Apache (license lock), receiptagent knobs

train_chaski.py

@@ -11,15 +11,9 @@
 #     "trackio",
 # ]
 # ///
-"""Chaski training — receiptagent pattern, not generic train_szl.py.
-
-LOCKED (do not invent):
-- Loss is TRL CE + Unsloth train_on_responses_only. No Λ / locked-8 / loop-tax term.
-- Λ uniqueness is Conjecture 1 (identity curriculum only).
-- Doctrine v11 LOCKED 749/14/163 @ c7c0ba17 (identity curriculum only).
-- seed 11, lr 2e-4, adamw_8bit, wd 0.01, LoRA r=16 α=16 (1.5B card).
-- Brain/A11oy corpus excluded.
-- Evals only if a held-out gate actually runs. This job records training MEASURED; eval is none-this-run.
+"""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).
 """
 from __future__ import annotations
 
@@ -34,14 +28,15 @@ from unsloth import FastLanguageModel
 from unsloth.chat_templates import train_on_responses_only
 from trl import SFTConfig, SFTTrainer
 
-MAX_SEQ_LEN = 1024
-BASE = os.environ.get("BASE_MODEL", "unsloth/Qwen2.5-1.5B-Instruct-bnb-4bit")
-CANONICAL_BASE = "Qwen/Qwen2.5-1.5B-Instruct"
+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"
 SEED = 11
 LORA_R = 16
-LORA_ALPHA = 16
+LORA_ALPHA = 32
+MAX_STEPS = 64
 
 OUROBOROS = [
     {
@@ -115,8 +110,8 @@ trainer = SFTTrainer(
     max_seq_length=MAX_SEQ_LEN,
     args=SFTConfig(
         per_device_train_batch_size=1,
-        gradient_accumulation_steps=8,
-        num_train_epochs=3,
+        gradient_accumulation_steps=2,
+        max_steps=MAX_STEPS,
         warmup_steps=6,
         learning_rate=2e-4,
         logging_steps=1,
@@ -127,7 +122,7 @@ trainer = SFTTrainer(
         output_dir="outputs",
         report_to="trackio",
         project="szl-chaski",
-        run_name="chaski-doctrine-sft-v1",
+        run_name="chaski-qwen35-0.8b-doctrine-v1",
         push_to_hub=True,
         hub_model_id=HUB,
         hub_private_repo=False,
@@ -150,7 +145,10 @@ print(f"[chaski] train done loss={loss} metrics={metrics}")
 adapter_dir = "chaski-adapter"
 model.save_pretrained(adapter_dir)
 tokenizer.save_pretrained(adapter_dir)
-model.save_pretrained_merged("chaski-merged", tokenizer, save_method="merged_16bit")
+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()
 receipt = {
@@ -158,13 +156,14 @@ receipt = {
     "schema": "szl.frontier-training-run/v1",
     "artifact": HUB,
     "base_model": CANONICAL_BASE,
+    "base_model_relation": "adapter",
     "base_model_runtime": BASE,
     "dataset": DATASET,
     "dataset_sha256": doctrine_sha,
     "extra_identity_turns": len(OUROBOROS),
     "training_rows": len(rows),
     "seed": SEED,
-    "num_train_epochs": 3,
+    "max_steps": MAX_STEPS,
     "warmup_steps": 6,
     "lora_r": LORA_R,
     "lora_alpha": LORA_ALPHA,