Qwen3-4B-Evasion

A fine-tuned model for detecting evasion levels in earnings call Q&A responses.

Model Description

Qwen3-4B-Evasion is a specialized model fine-tuned from Qwen/Qwen3-4B-Instruct-2507 for analyzing executive responses during earnings call Q&A sessions. The model classifies responses into three evasion categories based on the Rasiah taxonomy.

Intended Use

Primary Use Case

• Analyze transparency and directness of executive responses in earnings calls
• Financial discourse analysis
• Corporate communication research

Classification Categories

• direct: Clear, on-topic resolution to the question
• intermediate: Partially responsive, incomplete, or softened answer
• fully_evasive: Does not provide requested information

Training Details

Training Data

• Dataset: 27,097 earnings call Q&A pairs
• Source: Annotated by DeepSeek-V3.2 and Qwen3-Max models
• Label Distribution:
  • intermediate: 45.4%
  • direct: 29.8%
  • fully_evasive: 24.9%

Training Configuration

• Base Model: Qwen/Qwen3-4B-Instruct-2507
• Training Type: Full parameter fine-tuning
• Hardware: 2x NVIDIA B200 GPUs
• Epochs: 2
• Batch Size: 32 (effective)
• Learning Rate: 2e-5
• Framework: MS-SWIFT

Performance

Evaluated on 297 human-annotated benchmark samples:

Metric	Score
Overall Accuracy	75.08%
Weighted F1	74.75%
Weighted Precision	77.56%
Weighted Recall	75.08%

Per-Class Performance

Class	Precision	Recall	F1-Score	Support
direct	86.67%	54.74%	67.10%	95
intermediate	63.12%	80.91%	70.92%	110
fully_evasive	85.42%	89.13%	87.23%	92

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "FutureMa/Qwen3-4B-Evasion"
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_name)

# Prepare input
question = "What are your revenue projections for next quarter?"
answer = "We don't provide specific guidance on that."

prompt = f"""You are a financial discourse analyst. Classify the evasion level of this executive response.

Question: {question}
Answer: {answer}

Return JSON: {{"rasiah":"direct|intermediate|fully_evasive","confidence":0.00}}"""

messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=128, temperature=1)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Limitations

• Direct Class Recall: Lower recall (54.74%) for direct responses - model tends to be conservative
• Domain Specific: Optimized for earnings call context, may not generalize to other domains
• English Only: Trained exclusively on English text
• Confidence Calibration: Model confidence scores may require further calibration

Bias and Ethical Considerations

• Training data derived from corporate earnings calls may reflect existing biases in financial communication
• Model should not be used as sole determinant for investment decisions
• Human oversight recommended for critical applications

Citation

@misc{qwen3-4b-evasion,
  author = {FutureMa},
  title = {Qwen3-4B-Evasion: Earnings Call Evasion Detection Model},
  year = {2025},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/FutureMa/Qwen3-4B-Evasion}}
}

License

Apache 2.0

Acknowledgments

• Base model: Qwen Team
• Training framework: MS-SWIFT
• Evasion taxonomy: Rasiah et al.

Contact

For questions or issues, please open an issue on the model repository.