---
license: apache-2.0
base_model: Qwen/Qwen3-4B
language:
- en
pipeline_tag: text-generation
tags:
- text-classification
- natural-language-inference
- legal
- reinforcement-learning
- grpo
- qwen3
---

# eXTC — ContractNLI (3-class legal NLI)

> **Anonymized artifact for a paper under double-blind review.**
> Author identity and institution will be revealed at camera-ready.

This is the final-stage checkpoint of **eXTC** (eXplainable Text Classifier) for
**3-way natural language inference** over non-disclosure-agreement (NDA) clauses,
from the [ContractNLI](https://stanfordnlp.github.io/contract-nli/) benchmark.

- **Input**: a contract clause paired with a hypothesis.
- **Label**: `entailment`, `contradiction`, or `not_mentioned`.
- **Output**: a free-text reasoning trace followed by a final `LABEL: <label>` line —
  the reasoning serves as a local, inspectable explanation of the prediction.

## eXTC pipeline

eXTC is a three-stage explainable classifier. This checkpoint is the output of all three stages:

```
Qwen3-4B (base)
  │
  ├─ Stage I — SOP Learning (structured prompt optimization)
  │     A natural-language rulebook (Standard Operating Procedure) is learned
  │     via a structured prompt-optimization algorithm; used only to ground the
  │     teacher in Stage II (not present at inference).
  │
  ├─ Stage II — SOP-Grounded Reasoning Distillation (R-SFT)
  │     Teacher: gpt-4.1-mini, prompted with <SOP, input>, rejection sampling
  │     (M=4 traces/example, keep first trace whose label is correct).
  │     Student: Qwen3-4B fine-tuned with LoRA (r=64, alpha=128, 2 epochs) on the
  │     accepted reasoning+label traces, with class-balanced upsampling.
  │
  └─ Stage III — Beyond SOP via RL (BD-GRPO)
        Balanced Dynamic GRPO: per-class oversampling, then drop zero-advantage
        (homogeneous-rollout) groups and keep a class-balanced batch of
        informative groups, with a binary label-correctness reward.
```

The released checkpoint is the one with the best validation macro-F1 over the
RL training trajectory, evaluated on the held-out test set under that selection.

## Test metrics

ContractNLI 3-class test set (n=2091), greedy decoding (temperature=0):

| Metric | Value |
|---|---|
| Balanced accuracy | 0.8824 |
| Macro F1 | 0.8494 |
| Accuracy | 0.8871 |
| Invalid output rate | 0.001 |

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

repo = "extc-anon/extc-contractnli"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, dtype=torch.bfloat16, device_map="auto")

prompt = (
    "Premise: The Receiving Party shall not disclose Confidential Information to "
    "any third party without prior written consent.\n"
    "Hypothesis: The Receiving Party may share Confidential Information with its "
    "external auditors without consent.\n\n"
    "Classify the hypothesis as entailment, contradiction, or not_mentioned. "
    "Provide your reasoning and then the label."
)
text = tok.apply_chat_template(
    [{"role": "user", "content": prompt}],
    add_generation_prompt=True, tokenize=False,
)
ids = tok(text, return_tensors="pt").input_ids.to(model.device)
out = model.generate(ids, max_new_tokens=1024, do_sample=False)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))
```

## Format

- Standard HuggingFace `transformers` (safetensors, bfloat16, ~7.5 GB).
- Architecture: `Qwen3ForCausalLM`, 4.02B parameters.
- Test numbers above use greedy decoding (`do_sample=False`).

## License

Apache 2.0 (matches the [Qwen3 base model](https://huggingface.co/Qwen/Qwen3-4B)).

## Citation

Anonymous paper citation will be added at camera-ready.