Qwen3-8B-UnBias-Plus-SFT-Instruct-V2/README.md

---
license: apache-2.0
base_model: Qwen/Qwen3-8B
datasets:
- vector-institute/unbias-plus-dataset
language:
- en
tags:
- qwen3
- 8b
- bias-detection
- news-debiasing
- fine-tuned
- unsloth
- sft
- conversational
pipeline_tag: text-generation
---

# Qwen3-8B-UnBias-Plus-SFT-Instruct-V2

A fine-tuned version of [Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B) for news media bias detection and neutral rewriting, developed by the [Vector Institute](https://vectorinstitute.ai) as part of the [UnBias-Plus](https://github.com/VectorInstitute/unbias-plus) project.

Given a news article, the model identifies biased language segments, classifies each segment's bias type and severity, provides neutral replacements, assigns a single article-level severity score, and returns a fully rewritten unbiased version, all in a single structured JSON response.

This is the **V2 Instruct variant**, trained without chain-of-thought thinking blocks (`enable_thinking=False`). It produces clean structured JSON directly, making it suitable for production inference via vLLM or other OpenAI-compatible backends.

V2 is trained as a **conservative span-level bias annotator**: it flags only clear, material bias (while always flagging toxic/dehumanizing language), applies a strict bias-type label precedence, treats attributed/quoted language under the same standard, and changes only flagged text when producing the rewrite.

> **Trained on:** [train_4](https://huggingface.co/datasets/vector-institute/unbias-plus-dataset) (5,000 expert-annotated news articles)

## Difference from Qwen3-8B-UnBias-Plus-SFT-Instruct

|                          | SFT-Instruct (predecessor)                                   | SFT-Instruct-V2 (this model)                                                    |
| ------------------------ | ------------------------------------------------------------ | ------------------------------------------------------------------------------- |
| Training data            | train_3 (UnBias-Plus-3000)                                   | train_4 (5,000 expert-annotated articles)                                       |
| Article severity scale   | `0, 1, 2, 3, 4`                                              | `0–10` integer                                                                  |
| Segment severity         | `low` / `medium` / `high`                                    | `Low` / `Medium` / `High`                                                       |
| Output fields            | `severity`, `bias_found`, `biased_segments`, `unbiased_text` | `severity`, `biased_segments`, `unbiased_text` (label/bias_found derived downstream) |
| Bias type taxonomy       | 7 free-form labels                                           | 8 canonical slugs with strict label precedence                                  |
| System prompt            | Bias detection + rewrite rules                               | Conservative span-level annotator (label precedence, attribution handling, toxicity rules, audit step) |

## Model Details

| Property           | Value                                                    |
| ------------------ | -------------------------------------------------------- |
| Base model         | Qwen/Qwen3-8B                                            |
| Fine-tuning method | Supervised Fine-Tuning (SFT) with LoRA                   |
| Training precision | bf16 (full precision, no quantization during training)   |
| LoRA rank          | 16                                                       |
| Training framework | Unsloth + TRL                                            |
| Context length     | 8192 tokens                                              |
| Thinking mode      | Disabled (`enable_thinking=False`)                       |
| Output format      | Structured JSON                                          |
| Training dataset   | [train_4](https://huggingface.co/datasets/vector-institute/unbias-plus-dataset) (5,000 expert-annotated articles) |

## Usage

The exact system prompt used during training is maintained in
[`src/unbias_plus/prompt.py`](https://github.com/VectorInstitute/unbias-plus/blob/main/src/unbias_plus/prompt.py)
(`SYSTEM_PROMPT`). For faithful results, use that prompt verbatim — a condensed
version is shown below for illustration. The easiest path is to use the
[`unbias-plus`](https://github.com/VectorInstitute/unbias-plus) package, which
wires up the prompt, parsing, and offset computation for you.

### Using with transformers

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

model_id = "vector-institute/Qwen3-8B-UnBias-Plus-SFT-Instruct-V2"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
model.eval()

# Use the full SYSTEM_PROMPT from src/unbias_plus/prompt.py for best results.
SYSTEM_PROMPT = """You are a conservative span-level bias annotator.

Given an article, identify material biased language, assign one article-level
severity score (0-10), and produce a neutral rewrite that changes only flagged
text. Return the result only as a single valid JSON object with the fields
`severity`, `biased_segments`, and `unbiased_text`. Do not output `binary_label`
or `bias_found`. Do not write anything outside the JSON object."""

article = "Your news article here..."

messages = [
    {"role": "system", "content": SYSTEM_PROMPT},
    {"role": "user", "content": f"Analyze the following article for bias and return the result in the required JSON format.\n\nARTICLE:\n{article}"},
]

inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    enable_thinking=False,
    return_tensors="pt",
    return_dict=True,
    truncation=True,
    max_length=8192,
)

with torch.no_grad():
    outputs = model.generate(
        input_ids=inputs["input_ids"].to(model.device),
        attention_mask=inputs["attention_mask"].to(model.device),
        max_new_tokens=4096,
        do_sample=False,
        temperature=None,
        top_p=None,
        pad_token_id=tokenizer.eos_token_id,
    )

new_tokens = outputs[0][inputs["input_ids"].shape[1]:]
response = tokenizer.decode(new_tokens, skip_special_tokens=True)
result = json.loads(response)
```

### Using with vLLM

```bash
vllm serve vector-institute/Qwen3-8B-UnBias-Plus-SFT-Instruct-V2 \
  --served-model-name unbias-plus \
  --max-model-len 8192 \
  --dtype bfloat16
```

```python
from openai import OpenAI
import json

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")

completion = client.chat.completions.create(
    model="unbias-plus",
    messages=messages,
    max_tokens=4096,
    temperature=0,
    extra_body={"chat_template_kwargs": {"enable_thinking": False}},
)
result = json.loads(completion.choices[0].message.content)
```

## Output Schema

The model emits exactly three fields. `binary_label` and `bias_found` are **not**
produced by the model — they are derived downstream from `severity`
(`severity > 0` ⇒ biased).

```json
{
  "severity": 0,
  "biased_segments": [
    {
      "original": "exact substring from input article",
      "replacement": "neutral alternative phrase (or empty string to delete)",
      "severity": "Low | Medium | High",
      "bias_type": "loaded_language | euphemism | dehumanizing_language | opinion_as_fact | unsupported_generalization | stereotypical_association | sensationalism | informational_bias",
      "reasoning": "short, specific explanation of the linguistic cue"
    }
  ],
  "unbiased_text": "Full rewritten neutral article"
}
```

- `severity` is an integer `0–10` (article-level).
- Each segment's `severity` is a string: `Low`, `Medium`, or `High`.
- `bias_type` is exactly one of the eight canonical slugs listed below.
- If `biased_segments` is empty, article `severity` is `0`. If `severity > 0`, there is at least one segment.

### Article-level Severity Scale

| Value  | Meaning                                      |
| ------ | -------------------------------------------- |
| 0      | No biased segments                           |
| 1–5    | Limited, low, or moderate bias               |
| 6–10   | Strong, recurring, or highly distorting bias |

## Bias Types Detected

V2 assigns exactly one primary `bias_type` per segment, using the following
label precedence (first applicable label wins):

1. **`dehumanizing_language`** — treats people or groups as less than human, as a threatening mass, or as inherently dangerous (e.g. "vermin", "flood of migrants").
2. **`stereotypical_association`** — assigns traits, roles, or motives to a demographic/social/political/protected group, links identity to suspicion or wrongdoing, or reduces a group to a single function.
3. **`sensationalism`** — hyperbolic, alarmist, or dramatic language that inflates significance (e.g. "bombshell", "catastrophe", "sparks fury").
4. **`opinion_as_fact`** — an unattributed subjective or evaluative judgment presented as fact (e.g. "the policy is a failure").
5. **`unsupported_generalization`** — a sweeping or absolute claim about a group or situation without support (e.g. "everyone knows", "immigrants always").
6. **`euphemism`** — softened or vague wording that minimizes, hides, or downplays harmful facts (e.g. "collateral damage", "enhanced interrogation").
7. **`informational_bias`** — authorial framing that visibly treats one side, source, or claim differently in a way that steers interpretation.
8. **`loaded_language`** — clearly charged, morally weighted, editorial, toxic, abusive, or inflammatory wording not covered by a more specific type (e.g. "thugs", "so-called reform").


## Hardware Requirements

| Setup                | Configuration                                             |
| -------------------- | --------------------------------------------------------- |
| Recommended (server) | `torch_dtype=torch.bfloat16, device_map="auto"` (~16GB VRAM) |
| Lightweight (laptop) | `load_in_4bit=True` (~5GB VRAM)                           |

## Model Variants

| Model | Params | Context | Best for |
| ----- | ------ | ------- | -------- |
| [Qwen3.5-4B-UnBias-Plus-SFT-Instruct](https://huggingface.co/vector-institute/Qwen3.5-4B-UnBias-Plus-SFT-Instruct) | 4B | 4096 | Speed, low VRAM |
| Qwen3-8B-UnBias-Plus-SFT-Instruct-V2 (this) | 8B | 8192 | Higher recall, longer articles, production deployment |

## Citation

```@article{radwan2026unbias,
  title={UnBias-Plus: Detect, Explain, and Rewrite Bias},
  author={Radwan, Ahmed Y and ElKady, Ahmed and Chaduvula, Sindhuja and Hafez, Mohamed and Krishnan, Amrit and Raza, Shaina},
  journal={arXiv preprint arXiv:2606.23412},
  year={2026}
}
```

## Links

- 📊 Leaderboard: [UnBias-Plus Leaderboard](https://huggingface.co/spaces/vector-institute/UnBias-Plus-Leaderboard)
- 📁 Dataset: [vector-institute/unbias-plus-dataset](https://huggingface.co/datasets/vector-institute/unbias-plus-dataset)
- 🏛️ Organization: [Vector Institute](https://vectorinstitute.ai)
- 🌐 Project: [AIXpert](https://aixpert-project.eu/)