---
title: C2LLM-7B
canonical_url: "https://www.modelscope.ai/models/codefuse-ai/C2LLM-7B"
md_url: "https://www.modelscope.ai/models/codefuse-ai/C2LLM-7B.md"
repository: codefuse-ai/C2LLM-7B
last_updated: 2026-05-25
license: apache-2.0
pipeline_tag: feature-extraction
tasks:
  - feature-extraction
model_type:
  - c2llm
architectures:
  - C2LLMForEmbedding
base_model:
  - Qwen/Qwen2.5-Coder-7B-Instruct
base_model_relation: finetune
parameters: 7.7B
tensor_type:
  - BF16
library_name:
  - pytorch
  - transformer
  - sentence-transformers
  - safetensors
frameworks:
  - pytorch
language:
  - zh
  - en
downloads: 169
stars: 0
tags:
  - code
  - sentence-transformers
---

# C2LLM-7B

> C2LLM-7B - An open-source model by codefuse-ai on ModelScope. A New Frontier in Code Retrieval via Adaptive Cross-Attention Pooling

codefuse-ai/C2LLM-7B is a 7.7B-parameter feature-extraction model on ModelScope. licensed under apache-2.0. derived from Qwen/Qwen2.5-Coder-7B-Instruct.

- **Repository**: codefuse-ai/C2LLM-7B
- **License**: apache-2.0
- **Tasks**: feature-extraction
- **Parameters**: 7.7B
- **Base model**: Qwen/Qwen2.5-Coder-7B-Instruct
- **Tags**: code, sentence-transformers
- **Downloads**: 169
- **Stars**: 0
- **Last updated**: 2026-05-25

Source: https://www.modelscope.ai/models/codefuse-ai/C2LLM-7B

---

<div align="center" style="display: flex; justify-content: center; align-items: center; gap: 20px;">
    <a href="https://github.com/codefuse-ai/CodeFuse-Embeddings/tree/main/" style="display: flex; align-items: center; text-decoration: none; color: inherit;">
        <img src="https://github.githubassets.com/images/modules/logos_page/GitHub-Mark.png" width="30" height="30" style="vertical-align: middle; margin-right: 8px;">
        <span style="font-size: 1.5em; font-weight: bold;">CodeFuse-Embeddings</span>
    </a>
</div>



# A New Frontier in Code Retrieval via Adaptive Cross-Attention Pooling

[Paper](https://huggingface.co/papers/2512.21332) | [Code](https://github.com/codefuse-ai/CodeFuse-Embeddings)

**C2LLMs (Code Contrastive Large Language Models)** are powerful new models for generating code embeddings, designed to capture the deep semantics of source code. 

#### Key Features

- **Powerful Base Model**: Built upon the state-of-the-art `Qwen2.5-Coder`, inheriting its exceptional code comprehension capabilities.
- **Intelligent Pooling with PMA**: Instead of traditional `mean pooling` or `last token pooling`, C2LLM uses **PMA (Pooling by Multi-head Attention)**. This allows the model to dynamically focus on the most critical parts of the code, creating a more informative and robust embedding.
- **Trained for Retrieval**: C2LLM was fine-tuned on a massive dataset of **3 million query-document pairs**, optimizing it for real-world code retrieval and semantic search tasks. Supporting Text2Code/Code2Code/Code2Text tasks.

C2LLM is designed to be a go-to model for tasks like code search and Retrieval-Augmented Generation (RAG). For more details, please see our [GitHub repository](https://github.com/codefuse-ai/CodeFuse-Embeddings/tree/main). 

#  Model Details

# How to use

## Usage (**HuggingFace Transformers**)

```Python
from transformers import AutoModel, AutoTokenizer
import torch

model_path = "codefuse-ai/C2LLM-7B"

# Load the model
model = AutoModel.from_pretrained(model_path, torch_dtype=torch.bfloat16, trust_remote_code=True)

# Prepare your custom instruction
instruction = "xxxxx"

# Prepare the data
sentences = ['''int r = (int) params >> 8 & 0xff;
int p = (int) params & 0xff;

byte[] derived1 = SCrypt.scrypt(passwd.getBytes("UTF-8"), salt, N, r, p, 32);

if (derived0.length != derived1.length) return false;

int result = 0;
for (int i = 0; i < derived0.length; i++) {
result |= derived0[i] ^ derived1[i];
}
return result == 0;
} catch (UnsupportedEncodingException e) {
throw new IllegalStateException("JVM doesn't support UTF-8?");
} catch (GeneralSecurityException e) {
throw new IllegalStateException("JVM doesn't support SHA1PRNG or HMAC_SHA256?");
}
}''',
'''	
}
if (tempFrom > tempTo) {
return new RangeInfo(inclusive ? tempTo : tempTo + 1, tempFrom + 1, true);
}
return new RangeInfo(tempFrom, inclusive ? tempTo + 1 : tempTo, false);
}''']

sentences = [instruction+sentence for sentence in sentences]

# Get the embeddings
embeddings = model.encode(sentences)
```

## Usage (**Sentence-Transformers**)

```python
from sentence_transformers import SentenceTransformer

# Load the model
model = SentenceTransformer("codefuse-ai/C2LLM-7B", trust_remote_code=True, tokenizer_kwargs={"padding_side":"left"})

# Prepare your custom instruction
instruction = "xxxxx"

# Prepare the data
sentences = ['''int r = (int) params >> 8 & 0xff;
int p = (int) params & 0xff;

byte[] derived1 = SCrypt.scrypt(passwd.getBytes("UTF-8"), salt, N, r, p, 32);

if (derived0.length != derived1.length) return false;

int result = 0;
for (int i = 0; i < derived0.length; i++) {
result |= derived0[i] ^ derived1[i];
}
return result == 0;
} catch (UnsupportedEncodingException e) {
throw new IllegalStateException("JVM doesn't support UTF-8?");
} catch (GeneralSecurityException e) {
throw new IllegalStateException("JVM doesn't support SHA1PRNG or HMAC_SHA256?");
}
}''',
'''	
}
if (tempFrom > tempTo) {
return new RangeInfo(inclusive ? tempTo : tempTo + 1, tempFrom + 1, true);
}
return new RangeInfo(tempFrom, inclusive ? tempTo + 1 : tempTo, false);
}''']

sentences = [instruction+sentence for sentence in sentences]

# Get the embeddings
embeddings = model.encode(sentences)
```

## Evaluation (**MTEB**)

```python
from sentence_transformers import SentenceTransformer
from mteb.models import ModelMeta
from mteb.cache import ResultCache

model_name = "codefuse-ai/C2LLM-7B"

# Load the model
model = mteb.get_model(model_name) # if the model is not implemented in MTEB it will be eq. to SentenceTransformer(model_name)

# Select tasks
tasks = mteb.get_tasks(tasks=["AppsRetrieval", "CodeSearchNetCCRetrieval", "CodeEditSearchRetrieval","CodeSearchNetRetrieval","CodeFeedbackMT","CodeFeedbackST","CodeTransOceanContest","CodeTransOceanDL","COIRCodeSearchNetRetrieval","CosQA","StackOverflowQA","SyntheticText2SQL"])

# Cache the result
cache = ResultCache("./c2llm_results")

# Evaluate
results = mteb.evaluate(model, tasks=tasks, cache=cache, encode_kwargs={"batch_size": 16})
```

## Support Us

If you find this project helpful, please give it a star. It means a lot to us!

[![GitHub stars](https://img.shields.io/github/stars/codefuse-ai/CodeFuse-Embeddings?style=social)](https://github.com/codefuse-ai/CodeFuse-Embeddings/tree/main)

## Citation

@article{2025C2LLM,
  title={C2LLM Technical Report: A New Frontier in Code Retrieval via Adaptive Cross-Attention Pooling}, 
      author={Jin Qin and Zihan Liao and Ziyin Zhang and Hang Yu and Peng Di and Rui Wang},
  journal      = {CoRR},
  volume       = {abs/2512.21332},
  year         = {2025},
  url          = {https://doi.org/10.48550/arXiv.2512.21332},
  doi          = {10.48550/ARXIV.2512.21332},
  eprinttype    = {arXiv},
  eprint       = {2512.21332}
}