Eva-4B-V2

Eva-4B-V2

<p align="center"> <a href="https://huggingface.co/FutureMa/Eva-4B-V2"><img src="https://img.shields.io/badge/🤗-Model-yellow?style=for-the-badge" alt="Model"></a> <a href="https://huggingface.co/datasets/FutureMa/EvasionBench"><img src="https://img.shields.io/badge/🤗-Dataset-orange?style=for-the-badge" alt="Dataset"></a> <a href="https://github.com/IIIIQIIII/EvasionBench"><img src="https://img.shields.io/badge/GitHub-Repo-blue?style=for-the-badge" alt="GitHub"></a> <a href="https://iiiiqiiii.github.io/EvasionBench"><img src="https://img.shields.io/badge/Project-Page-green?style=for-the-badge" alt="Project Page"></a> <a href="https://colab.research.google.com/github/IIIIQIIII/EvasionBench/blob/main/scripts/eva4b_inference.ipynb"><img src="https://img.shields.io/badge/Colab-Quick_Start-F9AB00?style=for-the-badge&logo=googlecolab" alt="Open In Colab"></a> <a href="https://arxiv.org/abs/2601.09142"><img src="https://img.shields.io/badge/arXiv-Paper-red?style=for-the-badge" alt="Paper"></a> </p> <p align="center"> <b>A 4B parameter model fine-tuned for detecting evasive answers in earnings call Q&A sessions.</b> </p>

Model Description

• Base Model: Qwen/Qwen3-4B-Instruct-2507
• Task: Text Classification (Evasion Detection)
• Language: English
• License: Apache 2.0

Performance

Eva-4B-V2 achieves 84.9% Macro-F1 on the EvasionBench evaluation set, outperforming frontier LLMs:

<p align="center"> <img src="top5_performance.svg" alt="Top 5 Model Performance" width="100%"> </p>

Per-Class Performance

Label Definitions

Training

Two-Stage Training Pipeline

Qwen3-4B-Instruct-2507
        │
        ▼ Stage 1: 60K consensus data
        │
Eva-4B-Consensus
        │
        ▼ Stage 2: 24K three-judge data
        │
Eva-4B-V2

Training Configuration

Hardware

• Stage 1: 2x NVIDIA B200 (180GB SXM6)
• Stage 2: 4x NVIDIA H100 (80GB SXM5)

Usage

With Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_name = "FutureMa/Eva-4B-V2"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16, device_map="auto")

# Prompt template
prompt = """You are a financial analyst. Your task is to Detect Evasive Answers in Financial Q&A

Question: What is the expected margin for Q4?
Answer: We expect it to be 32%.

Response format:
```json
{"label": "direct|intermediate|fully_evasive"}
```

Answer in ```json content, no other text"""

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

with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=64, temperature=0.1, do_sample=False)

generated = outputs[0][inputs["input_ids"].shape[1]:]
print(tokenizer.decode(generated, skip_special_tokens=True))
# Output: ```json
# {"label": "direct"}
# ```

With vLLM

from vllm import LLM, SamplingParams

llm = LLM(model="FutureMa/Eva-4B-V2")
sampling_params = SamplingParams(temperature=0, max_tokens=64)

outputs = llm.generate([prompt], sampling_params)
print(outputs[0].outputs[0].text)

Links

Citation

@misc{ma2026evasionbenchlargescalebenchmarkdetecting,
  title={EvasionBench: A Large-Scale Benchmark for Detecting Managerial Evasion in Earnings Call Q&A},
  author={Shijian Ma and Yan Lin and Yi Yang},
  year={2026},
  eprint={2601.09142},
  archivePrefix={arXiv},
  primaryClass={cs.LG},
  url={https://arxiv.org/abs/2601.09142}
}

License

Apache 2.0