Qwen3.5-4B-UnBias-Plus-SFT-Instruct

A fine-tuned version of Qwen3.5-4B for news media bias detection and neutral rewriting, developed by the Vector Institute as part of the UnBias-Plus project.

Given a news article, the model identifies biased language segments, classifies their bias type and severity, provides neutral replacements, and returns a fully rewritten unbiased version — all in a single structured JSON response.

This is the Instruct variant — trained without chain-of-thought thinking blocks (enable_thinking=False). It is the fastest model in the UnBias-Plus family, with near-perfect rewrite length preservation and zero duplicate segment rate.

Trained on: train_3 — UnBias-Plus-3000 (3,000 expert-annotated news articles)

Model Details

Property	Value
Base model	Qwen/Qwen3.5-4B
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	4096 tokens
Thinking mode	Disabled (`enable_thinking=False`)
Output format	Structured JSON
Training dataset	UnBias-Plus-3000 (train_3)

Usage

Using with transformers

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch, json

model_id = "vector-institute/Qwen3.5-4B-UnBias-Plus-SFT-Instruct"

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

SYSTEM_PROMPT = """You are an expert linguist and bias detection specialist.
Your task is to carefully read a news article, detect ALL biased language,
and return a structured JSON response. Return ONLY valid JSON, no extra text."""

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=4096,
)

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=2048,
        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

vllm serve vector-institute/Qwen3.5-4B-UnBias-Plus-SFT-Instruct \
  --max-model-len 4096

from openai import OpenAI
import json

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

completion = client.chat.completions.create(
    model="vector-institute/Qwen3.5-4B-UnBias-Plus-SFT-Instruct",
    messages=messages,
    max_tokens=2048,
    temperature=0,
    extra_body={"chat_template_kwargs": {"enable_thinking": False}},
)
result = json.loads(completion.choices[0].message.content)

Output Schema

{
  "severity": 0 | 1 | 2 | 3 | 4,
  "bias_found": true | false,
  "biased_segments": [
    {
      "original": "exact substring from input article",
      "replacement": "neutral alternative phrase",
      "severity": "high" | "medium" | "low",
      "bias_type": "loaded language | dehumanizing framing | false generalizations | framing bias | euphemism/dysphemism | politically charged terminology | sensationalism",
      "reasoning": "1-2 sentence explanation"
    }
  ],
  "unbiased_text": "Full rewritten neutral article"
}

Severity Scale

Value	Meaning
0	Neutral — no bias detected
1	Minor isolated loaded words
2	Recurring biased framing
3	Strong persuasive tone
4	Inflammatory rhetoric

Bias Types Detected

Loaded language — words with strong emotional connotations
Dehumanizing framing — language that strips dignity from groups
False generalizations — sweeping statements ("they always", "all of them")
Framing bias — selective wording that implies a viewpoint
Euphemism/dysphemism — softening or hardening language to manipulate perception
Politically charged terminology — labels used to provoke rather than describe
Sensationalism — exaggerated language to evoke emotional responses

Evaluation Results

Evaluated on BABE-350 (175 biased + 175 unbiased, balanced), judge: GPT-4o-mini.

Metric	Score
Parse rate	99.1%
Bias reduction %	57.5%
Contextual relevance	4.09 / 5
Global rewrite quality	3.15 / 5
ROUGE-L vs human reference	0.709
Recall at biased words	0.847
Segment replacement quality	3.88 / 5
Duplicate rate	0.0%
Length ratio (median)	1.000
Latency (median)	19.8s

See the full UnBias-Plus Leaderboard for model comparisons.

Hardware Requirements

Setup	Configuration
Recommended (server)	`torch_dtype=torch.bfloat16, device_map="auto"` (~9GB VRAM)
Lightweight (laptop)	`load_in_4bit=True` (~3GB VRAM)

Model Variants

Model	Params	Context	Best for
Qwen3.5-4B-UnBias-Plus-SFT-Instruct (this)	4B	4096	Speed, low VRAM, best segment quality
Qwen3-8B-UnBias-Plus-SFT-Instruct	8B	8192	Higher recall, longer articles
Qwen3-8B-UnBias-Plus-SFT-Instruct-Legacy	8B	8192	Best unbiased handling, deployed on GCP

Limitations

Trained primarily on English-language news articles
Unbiased article handling is weaker than the Legacy model (correct identification median 2.0 vs 5.0)
Context limited to 4096 tokens — may truncate very long articles
Outputs should be reviewed by a human before use in production

Citation

If you use this model in your research or application, please cite:

  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}
}

vector-institute/Qwen3.5-4B-UnBias-Plus-SFT-Instruct