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>

Dockerfile.train

@@ -12,36 +12,19 @@
 #       --dataset /data/datasets/routing_decision-v1.jsonl \
 #       --dataset /data/datasets/tlp_assignment-v1.jsonl \
 #       --output /data/adapters/psyc-v1
+#
+# Base image already ships Python 3.11 + torch 2.6 + CUDA 12.4 + cuDNN9, so
+# there is no apt step and no torch download. Qwen3.5 needs transformers v5 —
+# unsloth pulls it automatically.
 
-FROM nvidia/cuda:12.4.1-cudnn-devel-ubuntu22.04
+FROM pytorch/pytorch:2.6.0-cuda12.4-cudnn9-devel
 
-ENV DEBIAN_FRONTEND=noninteractive \
-    PYTHONUNBUFFERED=1 \
+ENV PYTHONUNBUFFERED=1 \
     PIP_NO_CACHE_DIR=1 \
     HF_HOME=/data/.hf-cache
 
-RUN apt-get update && apt-get install -y --no-install-recommends \
-        python3.11 python3.11-venv python3-pip \
-        git curl ca-certificates \
-    && rm -rf /var/lib/apt/lists/* \
-    && ln -sf /usr/bin/python3.11 /usr/local/bin/python \
-    && ln -sf /usr/bin/python3.11 /usr/local/bin/python3
-
-RUN python -m pip install --upgrade pip wheel setuptools && \
-    python -m pip install \
-        torch==2.5.1 \
-        --index-url https://download.pytorch.org/whl/cu124
-
-RUN python -m pip install \
-        "unsloth @ git+https://github.com/unslothai/unsloth.git" \
-        transformers>=4.46 \
-        datasets>=3.0 \
-        peft>=0.13 \
-        trl>=0.12 \
-        accelerate>=1.1 \
-        bitsandbytes>=0.44 \
-        sentencepiece \
-        protobuf
+RUN pip install --upgrade pip && \
+    pip install unsloth unsloth_zoo trl datasets
 
 WORKDIR /workspace
 COPY scripts/train_qlora.py /workspace/train_qlora.py