Qwen2.5 7B Latent Verification

🚀 Qwen2.5-7B-Instruct

This is a special version of Qwen2.5-7B-Instruct with latent-space verification. It can detect and correct factual inaccuracies before they appear in the output, enhancing the factual consistency of the model.

🚀 Quick Start

Example Usage with Verification

from latent_verification import load_verification_model
from transformers import AutoTokenizer

# Load the verification-enhanced version
verified_model_name = "YourCustomOrg/Qwen2.5-7B-Instruct-Verification"
model = load_verification_model(verified_model_name)
tokenizer = AutoTokenizer.from_pretrained(verified_model_name)

prompt = "The capital of France is Marseilles, correct?"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.no_grad():
    outputs = model.generate(inputs["input_ids"], max_new_tokens=50)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

How to Add Verification to Your Own Model

from transformers import AutoModelForCausalLM
from latent_verification import create_verification_model

base_model_name = "Qwen/Qwen2.5-7B-Instruct"
base_model = AutoModelForCausalLM.from_pretrained(base_model_name)

# Add latent verification adapters
verified_model = create_verification_model(
    base_model=base_model,
    adapter_locations=[2, 5, 8, 11, 14, 17, 20, 27],  # Example: places adapters in these layers
    bottleneck_size=64,
    enable_cross_layer=True
)

# Fine-tune only the verification parameters (base model remains frozen)
# (See the repository's training scripts for full details)
verified_model.save_pretrained("YourCustomOrg/Qwen2.5-7B-Instruct-Verification")

✨ Features

Latent-Space Verification: Self-Correcting Implementation

This special version of Qwen2.5-7B-Instruct incorporates Latent-Space Verification based on the approach described in "Latent-Space Verification for Self-Correcting LLMs" (Warren, 2025). The verification mechanism embeds lightweight adapters (LoRA-style) into the hidden layers of the transformer to detect and correct factual inaccuracies before they emerge in the output.

Key Highlights

• Minimal Parameter Overhead: Less than 0.1% additional parameters (about 6.3M for a 7.6B model).
• Inside-the-Model Verification: The approach intercepts hidden states to detect/correct factual errors.
• Improved Accuracy: Achieves up to ~10% absolute gains in factual consistency on certain benchmarks.
• Architecture-Agnostic: Verification adapters can be placed in various model families with minimal changes.

Original Qwen2.5-7B-Instruct Features

Qwen2.5 is the latest series of Qwen large language models. For Qwen2.5, we release a number of base language models and instruction-tuned language models ranging from 0.5 to 72 billion parameters. Qwen2.5 brings the following improvements upon Qwen2:

• Significantly more knowledge and greatly improved capabilities in coding and mathematics, thanks to our specialized expert models in these domains.
• Significant improvements in instruction following, generating long texts (over 8K tokens), understanding structured data (e.g., tables), and generating structured outputs especially JSON. More resilient to the diversity of system prompts, enhancing role-play implementation and condition-setting for chatbots.
• Long-context Support up to 128K tokens and can generate up to 8K tokens.
• Multilingual support for over 29 languages, including Chinese, English, French, Spanish, Portuguese, German, Italian, Russian, Japanese, Korean, Vietnamese, Thai, Arabic, and more.

This repo contains the instruction-tuned 7B Qwen2.5 model, which has the following features:

📦 Installation

The code of Qwen2.5 is included in the latest Hugging Face transformers. We advise using the latest version of transformers.

With transformers<4.37.0, you will encounter the following error:

KeyError: 'qwen2'

📚 Documentation

Evaluation & Performance

Detailed evaluation results are in the Latent-Space Verification paper. For GPU memory and throughput benchmarks, see here.

The verification mechanism can improve factual reliability by ~10% in many tasks while preserving or even enhancing the base model’s fluency. In practice, the overall GPU footprint remains almost identical, with a small overhead for verification steps.

Processing Long Texts

The current config.json is set for a context length up to 32,768 tokens. To handle inputs exceeding 32,768 tokens, we use YaRN, a method for length extrapolation that preserves strong performance on long texts.

For supported frameworks, you can add this snippet to config.json to enable YaRN:

{
  ...,
  "rope_scaling": {
    "factor": 4.0,
    "original_max_position_embeddings": 32768,
    "type": "yarn"
  }
}

When deploying, we recommend vLLM. Please refer to our docs for usage details. Note that current vLLM only supports a static rope_scaling, which may affect shorter text performance if you enable very large factors.

More Information

For more information, please see our blog, GitHub, and Documentation.

🔧 Technical Details

Qwen2.5 Improvements

• Knowledge and Capabilities: Significantly more knowledge and greatly improved capabilities in coding and mathematics, thanks to our specialized expert models in these domains.
• Instruction Following and Text Generation: Significant improvements in instruction following, generating long texts (over 8K tokens), understanding structured data (e.g., tables), and generating structured outputs especially JSON. More resilient to the diversity of system prompts, enhancing role-play implementation and condition-setting for chatbots.
• Long-context Support: Long-context Support up to 128K tokens and can generate up to 8K tokens.
• Multilingual Support: Multilingual support for over 29 languages, including Chinese, English, French, Spanish, Portuguese, German, Italian, Russian, Japanese, Korean, Vietnamese, Thai, Arabic, and more.

Model Architecture

• Type: Causal Language Models
• Training Stage: Pretraining & Post-training
• Architecture: transformers with RoPE, SwiGLU, RMSNorm, and Attention QKV bias
• Number of Parameters: 7.61B
• Number of Parameters (Non-Embedding): 6.53B
• Number of Layers: 28
• Number of Attention Heads (GQA): 28 for Q and 4 for KV
• Context Length: Full 131,072 tokens and generation up to 8192 tokens

📄 License

This project is licensed under the Apache-2.0 License.

📖 Citation

If you find our work helpful, feel free to cite Qwen2.5 and Latent-Space Verification together.

Qwen2.5:

@misc{qwen2.5, title = {Qwen2.5: A Party of Foundation Models}, url = {https://qwenlm.github.io/blog/qwen2.5/}, author = {Qwen Team}, month = {September}, year = {2024} }

@article{qwen2, title={Qwen2 Technical Report}, author={An Yang and Baosong Yang and Binyuan Hui and et al.}, journal={arXiv preprint arXiv:2407.10671}, year={2024} }

Latent-Space Verification:

@misc{warren2025latent, title={Latent-Space Verification for Self-Correcting LLMs}, author={Warren, Jacob}, year={2025}, publisher={GitHub}, journal={GitHub repository}, howpublished={\url{https://github.com/jacobwarren/Latent-Space-Verification-for-Self-Correcting-LLMs}} }