README.md

---
base_model:
- Qwen/Qwen3-1.7B
datasets:
- codefuse-ai/F2LLM
language:
- en
license: apache-2.0
pipeline_tag: feature-extraction
library_name: transformers
---

# F2LLM Technical Report: Matching SOTA Embedding Performance with 6 Million Open-Source Data

This model is presented in the paper [F2LLM Technical Report: Matching SOTA Embedding Performance with 6 Million Open-Source Data](https://huggingface.co/papers/2510.02294).
The code for this model is available on [GitHub](https://github.com/codefuse-ai/CodeFuse-Embeddings/tree/main/F2LLM).

F2LLMs (Foundation to Feature Large Language Models) are foundation models directly finetuned on 6 million high-quality query-document pairs (available in [codefuse-ai/F2LLM](https://huggingface.co/datasets/codefuse-ai/F2LLM)) covering a diverse range of retrieval, classification, and clustering data, curated solely from open-source datasets without any synthetic data. These models are trained with homogeneous macro batches in a single stage, without sophisticated multi-stage pipelines.

## Usage

To encode a batch of sentences:

```python
from transformers import AutoModel, AutoTokenizer
import torch
import torch.nn.functional as F


model_path = "codefuse-ai/F2LLM-1.7B"
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModel.from_pretrained(model_path, torch_dtype=torch.bfloat16, device_map={'': 0})

sentences = [
    'We present F2LLM, a family of fully open embedding LLMs that achieve a strong balance between model size, training data, and embedding performance.',
    'Model checkpoints, training datasets, and training code are released, positioning F2LLM as a strong, reproducible, and budget-friendly baseline for future research in text embedding models.'
]

def encode(sentences):
    batch_size = len(sentences)
    sentences = [s+tokenizer.eos_token for s in sentences]
    tokenized_inputs = tokenizer(sentences, padding=True, return_tensors='pt', add_special_tokens=False).to(model.device)
    last_hidden_state = model(**tokenized_inputs).last_hidden_state
    eos_positions = tokenized_inputs.attention_mask.sum(dim=1) - 1
    embeddings = last_hidden_state[torch.arange(batch_size, device=model.device), eos_positions]
    embeddings = F.normalize(embeddings, p=2, dim=1)
    return embeddings

embeddings = encode(sentences)
```

## Evaluation

To evaluate F2LLMs on MTEB (currently requires installing MTEB from source):

```python
import mteb
import logging
logging.basicConfig(level=logging.INFO)

task_names = ['AmazonCounterfactualClassification', 'ArXivHierarchicalClusteringP2P', 'ArXivHierarchicalClusteringS2S', 'ArguAna', 'AskUbuntuDupQuestions', 'BIOSSES', 'Banking77Classification', 'BiorxivClusteringP2P.v2', 'CQADupstackGamingRetrieval', 'CQADupstackUnixRetrieval', 'ClimateFEVERHardNegatives', 'FEVERHardNegatives', 'FiQA2018', 'HotpotQAHardNegatives', 'ImdbClassification', 'MTOPDomainClassification', 'MassiveIntentClassification', 'MassiveScenarioClassification', 'MedrxivClusteringP2P.v2', 'MedrxivClusteringS2S.v2', 'SCIDOCS', 'SICK-R', 'STS12', 'STS13', 'STS14', 'STS15', 'STS17', 'STS22.v2', 'STSBenchmark', 'SprintDuplicateQuestions', 'StackExchangeClustering.v2', 'StackExchangeClusteringP2P.v2', 'SummEvalSummarization.v2', 'TRECCOVID', 'Touche2020Retrieval.v3', 'ToxicConversationsClassification', 'TweetSentimentExtractionClassification', 'TwentyNewsgroupsClustering.v2', 'TwitterSemEval2015', 'TwitterURLCorpus', 'MindSmallReranking']

tasks = [
    mteb.get_task(task_name, languages = ["eng"], eval_splits=["test"], exclusive_language_filter=True)
    for task_name in task_names
]


model = mteb.get_model("codefuse-ai/F2LLM-1.7B", device="cuda:0")
evaluation = mteb.MTEB(tasks=tasks)
evaluation.run(model, encode_kwargs={"batch_size": 16})
```

## Training

Training code is available in our [Github repo](https://github.com/codefuse-ai/CodeFuse-Embeddings/tree/main/F2LLM).

## Citation

If you use the F2LLM models, data, or code, please cite the following technical report.

```
@article{2025F2LLM,
  title={F2LLM Technical Report: Matching SOTA Embedding Performance with 6 Million Open-Source Data}, 
  author={Ziyin Zhang and Zihan Liao and Hang Yu and Peng Di and Rui Wang},
  journal      = {CoRR},
  volume       = {abs/2510.02294},
  year         = {2025},
  url          = {https://doi.org/10.48550/arXiv.2510.02294},
  doi          = {10.48550/ARXIV.2510.02294},
  eprinttype    = {arXiv},
  eprint       = {2510.02294}
}
```