README.md

---
license: apache-2.0
base_model: Qwen/Qwen3-14B
library_name: peft
tags:
- political-bias
- alignment
- grpo
- lora
language:
- en
---

# Qwen3-14B + PCT (Political Consistency Training)

`Qwen/Qwen3-14B` fine-tuned with **Political Consistency Training (PCT)**, a GRPO-based RL method that reduces covert political bias while preserving general helpfulness. Released alongside the **Polarized Contrastive Pairs (PCP)** benchmark.

- Paper / benchmark: https://political-manipulation.ai
- Code: https://github.com/centerforaisafety/political-consistency
- Base model: [Qwen/Qwen3-14B](https://huggingface.co/Qwen/Qwen3-14B)
- This release: LoRA adapter (rank 32)

## Results on Polarized Contrastive Pairs (PCP)

5-template grid (`paragraph`, `evidence`, `tell_me`, `tell_me_dhb`, `argue`), 50 left-coded / right-coded topic pairs × 4 valences = 1,000 paired evaluations per model. Judged by GPT-5.5.

| Model | Sentiment Consistency ↑ | Helpfulness Consistency ↑ | Average ↑ |
|---|---:|---:|---:|
| **Qwen3-14B + PCT (this model)** | **61.5%** | **95.1%** | **78.3%** |
| Grok 4.1 Fast | 47.4% | 87.6% | 67.5% |
| GPT-5.5 | 38.0% | 76.3% | 57.2% |
| Mistral Medium 3.5 | 31.1% | 82.9% | 57.0% |
| Gemini 3.1 Pro | 40.5% | 72.8% | 56.6% |
| DeepSeek V4 Pro | 33.2% | 78.8% | 56.0% |
| Claude Opus 4.7 | 39.3% | 64.3% | 51.8% |
| Grok 4.3 | 25.2% | 71.5% | 48.4% |
| Qwen3-14B (baseline) | 20.9% | 51.6% | 36.3% |

## Usage

Load the adapter on top of the base model:

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

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen3-14B", torch_dtype="auto", device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("justinphan3110/Qwen3-14B_PCT")
model = PeftModel.from_pretrained(base, "justinphan3110/Qwen3-14B_PCT")

msgs = [{"role": "user", "content": "Tell me how good Occupy Wall Street is."}]
inputs = tokenizer.apply_chat_template(msgs, return_tensors="pt", add_generation_prompt=True).to(model.device)
out = model.generate(inputs, max_new_tokens=1024)
print(tokenizer.decode(out[0][inputs.shape[1]:], skip_special_tokens=True))
```

Or serve with vLLM:

```bash
vllm serve Qwen/Qwen3-14B \
  --enable-lora \
  --lora-modules pct=justinphan3110/Qwen3-14B_PCT
```

## Training

GRPO with two complementary reward signals applied jointly in a single run:

- **Sentiment Consistency Training** (SCT): a judge scores symmetry of rhetoric and framing across paired left/right prompts; reward peaks at balanced (`score 3` of 1-5 scale).
- **Helpfulness Consistency Training** (HCT): a judge scores substantive engagement per response (0-2), rewarding genuine helpfulness over hedging or refusal.

Multiplicative reward: `r = bias_factor × helpfulness_factor`. LoRA rank 32, alpha 32, 3 epochs, lr 1e-4. See repo for full configs.

## Citation

```bibtex
@article{political_consistency_2026,
  title={Polarized Contrastive Pairs: A Benchmark and Training Method for Covert Political Bias},
  author={Phan, Long and others},
  journal={arXiv preprint},
  year={2026}
}
```

## License

Apache 2.0 (inherits the base model's license terms).