---
library_name: peft
base_model: Qwen/Qwen3.5-9B
license: apache-2.0
tags:
- peft
- lora
- qlora
- context-selection
- llm-agent
- decision-aware-context
---

# CICL Qwen3.5-9B QLoRA Adapter

This repository contains a PEFT LoRA adapter trained for decision-aware context judgment experiments in CICL. The repository does not include the Qwen3.5-9B base model.

It is an **adapter-only** release. To run it, load the adapter with the matching base model:

```text
base model: Qwen/Qwen3.5-9B
adapter:    XinyuGuan/CICL
```

The base model is not redistributed here and is governed by the base model provider's own terms.

## Contents

- `adapter_model.safetensors`: LoRA adapter weights.
- `adapter_config.json`: PEFT adapter configuration, with `Qwen/Qwen3.5-9B` as the base model reference.
- `tokenizer.json`, `tokenizer_config.json`, `chat_template.jinja`: tokenizer and chat-format files used during evaluation.
- `train_metrics.json`, `eval_field_report.json`, `selection_agreement_n20.json`: aggregate training and evaluation summaries.

Per-example teacher traces, API credentials, private prompts, and intermediate optimizer states are not included.

## Download

```bash
hf download XinyuGuan/CICL \
  --local-dir artifacts/hf_release/cicl-qwen35-qlora-adapter
```

## Intended Use

The adapter is intended for reproducing the CICL surrogate-judge experiments. It should be loaded with the matching Qwen3.5-9B base model through PEFT.

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "Qwen/Qwen3.5-9B"
adapter_id = "XinyuGuan/CICL"

tokenizer = AutoTokenizer.from_pretrained(adapter_id, trust_remote_code=True)
base = AutoModelForCausalLM.from_pretrained(
    base_id,
    trust_remote_code=True,
    device_map="auto",
)
model = PeftModel.from_pretrained(base, adapter_id)
```

## Use with the CICL Codebase

In the CICL repository, `qwen_local` defaults to the Hugging Face base model and this adapter:

```text
QWEN_LOCAL_BASE=Qwen/Qwen3.5-9B
QWEN_LOCAL_ADAPTER=XinyuGuan/CICL
```

Example preflight command:

```bash
python3 -m cicl_agent.evaluation.llm_agreement_preflight \
  --repo experiments/data/synthetic/v1/repo \
  --tasks experiments/data/synthetic/v1/tasks.jsonl \
  --teacher-examples artifacts/outputs/latest/synthetic_opus_v1/llm_examples.clean.jsonl \
  --llm-provider qwen_local
```

Example field-level evaluation:

```bash
PYTHONPATH=. CUDA_VISIBLE_DEVICES=0 python3 -m training.scripts.eval_qwen_judge \
  --base Qwen/Qwen3.5-9B \
  --adapter XinyuGuan/CICL \
  --val training/data/opus_v1/val.jsonl \
  --output artifacts/outputs/latest/qwen_local_eval/eval_field_report.json
```

## Evaluation Snapshot

On the held-out validation split used in the CICL experiments, the adapter produced parseable JSON for all 144 evaluated examples. Mean absolute error was below 0.07 across the five scalar judgment fields reported in `eval_field_report.json`.

These numbers are intended as diagnostic evidence for the paper's surrogate-judge study. They should not be interpreted as a general-purpose replacement for stronger teacher models.

## Limitations

- This is not a standalone language model; it requires `Qwen/Qwen3.5-9B`.
- It is trained for CICL counterfactual context-judgment experiments, not general chat or coding-agent use.
- It should not be described as equivalent to Claude/Opus teacher models.
- It is intended to support reproducibility of the surrogate-judge and selection-agreement experiments reported with the CICL project.