stage-3c: working QLoRA training + eval — pytorch base, Qwen3.5 slug, SFTConfig

Training and eval now run clean on the unsloth 2026.5.2 / transformers v5 /
torch 2.10 stack. Fixes: pytorch/pytorch base image (sidesteps the nvidia/cuda
apt-signature failure and the torch download), correct base-model slug
unsloth/Qwen3.5-4B, TRL SFTConfig API. Adds scripts/eval_adapter.py — runs
dataset rows through base+adapter with structured (transformers-v5) message
content and Qwen3.5 thinking-mode stripping.

First v1 adapter: loss 2.10 -> 0.32 over 3 epochs. Eval surfaced an ill-posed
ioc_extraction dataset (output URL not present in input) — to be fixed in the
ExampleBuilder before the next training run.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

scripts/train_qlora.py

@@ -22,8 +22,7 @@ from pathlib import Path
 from typing import Dict, List
 
 from datasets import Dataset
-from transformers import TrainingArguments
-from trl import SFTTrainer
+from trl import SFTConfig, SFTTrainer
 
 
 def load_examples(paths: List[Path]) -> List[Dict[str, str]]:
@@ -44,7 +43,7 @@ def load_examples(paths: List[Path]) -> List[Dict[str, str]]:
 def main() -> None:
     parser = argparse.ArgumentParser(description=__doc__)
     parser.add_argument("--dataset", action="append", required=True, help="JSONL path (repeatable)")
-    parser.add_argument("--base-model", default="unsloth/Qwen3.5-4B-Instruct-bnb-4bit")
+    parser.add_argument("--base-model", default="unsloth/Qwen3.5-4B")
     parser.add_argument("--output", default="/data/adapters/psyc-v1")
     parser.add_argument("--epochs", type=int, default=3)
     parser.add_argument("--lr", type=float, default=2e-4)
@@ -96,9 +95,9 @@ def main() -> None:
         model=model,
         tokenizer=tokenizer,
         train_dataset=dataset,
-        dataset_text_field="text",
-        max_seq_length=args.max_seq_length,
-        args=TrainingArguments(
+        args=SFTConfig(
+            dataset_text_field="text",
+            max_seq_length=args.max_seq_length,
             per_device_train_batch_size=args.batch_size,
             gradient_accumulation_steps=args.grad_accum,
             warmup_steps=5,