SkillRet-Embedding-0.6B

This is a sentence-transformers model fine-tuned for AI agent skill retrieval. Given a natural-language user request, the model retrieves relevant agent skills from a large skill library.

The model is fine-tuned from Qwen/Qwen3-Embedding-0.6B on the SkillRet benchmark training split using contrastive learning (MultipleNegativesRankingLoss).

📄 Technical report: SkillRet: A Large-Scale Benchmark for Skill Retrieval in LLM Agents (arXiv:2605.05726)

Usage

Sentence Transformers

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("ThakiCloud/SkillRet-Embedding-0.6B", trust_remote_code=True)

query_prompt = "Instruct: Given a skill search query, retrieve relevant skills that match the query\nQuery: "

queries = [
    query_prompt + "Help me set up a CI/CD pipeline for my Python project"
]
skills = [
    "ci-cd-setup | Configure continuous integration and deployment pipelines ...",
    "python-debugging | Debug Python applications using pdb and logging ...",
]

q_emb = model.encode(queries, normalize_embeddings=True)
s_emb = model.encode(skills, normalize_embeddings=True)

similarities = q_emb @ s_emb.T
print(similarities)

Training Details

Base model: Qwen3-Embedding-0.6B (0.6B parameters)
Training data: SkillRet benchmark training split (127,190 query–skill pairs from 63,259 queries and 10,123 skills)
Loss: MultipleNegativesRankingLoss (InfoNCE) with cross-GPU negative sharing
Hardware: 4× NVIDIA B200 GPUs (DDP)
Effective batch size: 384 (96 per device × 4 GPUs)
Max sequence length: 8,192 tokens
Learning rate: 2e-5
Epochs: 1
Training time: ~6 hours
Precision: BF16

Training Logs

Epoch	Step	Training Loss	NDCG@15
0.15	50	2.4288	0.7802
0.30	100	1.9920	0.7842
0.45	150	1.9758	0.7887
0.60	200	1.9011	0.7865
0.76	250	1.9100	0.7874
0.91	300	1.9412	0.7859
1.0	331	—	0.7862

Best checkpoint at step 150 (bold row).

Evaluation Results

Evaluated on the SkillRet benchmark test split (4,997 queries, 6,660 skills).

Metric	@5	@10	@15
NDCG	0.7557	0.7803	0.7887
Recall	0.7915	0.8542	0.8809
Completeness	0.6596	0.7509	0.7903

Intended Use

This model is designed for retrieving agent skills given natural-language user requests. It is part of the SkillRet benchmark submission for evaluating skill retrieval systems for AI agents.

Limitations

Optimized for English-language queries and agent skills.
Performance may vary on domains outside the SkillRet benchmark distribution.
The model retrieves skills but does not execute them.

Framework Versions

Python: 3.10.12
Sentence Transformers: 5.4.1
Transformers: 5.5.4
PyTorch: 2.7.1+cu128

Citation

If you use this model or the SkillRet benchmark, please cite:

@article{cho2026skillret,
  title   = {SkillRet: A Large-Scale Benchmark for Skill Retrieval in LLM Agents},
  author  = {Cho, Hongcheol and Kang, Ryangkyung and Kim, Youngeun},
  journal = {arXiv preprint arXiv:2605.05726},
  year    = {2026},
  url     = {https://arxiv.org/abs/2605.05726}
}

Paper: https://arxiv.org/abs/2605.05726