We introduce a tool named InsTag for analyzing supervised fine-tuning (SFT) data in LLM aligning with human preference. For local tagging deployment, we release InsTagger, fine-tuned on InsTag results, to tag the queries in SFT data. Through the scope of tags, we sample a 6K subset of open-resourced SFT data to fine-tune LLaMA and LLaMA-2 and the fine-tuned models TagLM-13B-v1.0 and TagLM-13B-v2.0 outperform many open-resourced LLMs on MT-Bench.

🤗 InsTagger Checkpoint • 👉 Online LocalTagger Demo • 📖 Paper

🤖️ TagLM-13B-v1.0 Checkpoint 🤖️ TagLM-13B-v2.0 Checkpoint

What is InsTag?

Foundation language models obtain the instruction-following ability through supervised fine-tuning (SFT). Diversity and complexity are considered critical factors of a successful SFT dataset, while their definitions remain obscure and lack quantitative analyses. In this work, we propose InsTag, an open-set fine-grained tagger, to tag samples within SFT datasets based on semantics and intentions and define instruction diversity and complexity regarding tags. We obtain 6.6K tags to describe comprehensive user queries. We analyze popular open-sourced SFT datasets and find that the model ability grows with more diverse and complex data. Based on this observation, we propose a data selector based on InsTag to select 6K diverse and complex samples from open-source datasets and fine-tune models on InsTag-selected data. These models outperform open-source models based on considerably larger SFT data evaluated by MT-Bench, echoing the importance of query diversity and complexity.

• [08/2023] 🔥 We have an online demo of InsTagger hosted by ModelScope. Please refer to the link on the top. Thanks ModelScope!

• [08/2023] 🔥 We released aligned LLMs TagLM-13B-v1.0 and TagLM-13B-v2.0 based on LLaMA and LLaMA-2 respectively. Both are fine-tuned on sub-sampled SFT data according to InsTag. Download v1.0 and v2.0.

• [08/2023] 🔥 We released an LLM InsTagger fine-tuned on our tagging results for local tagging deployments. Download weight.

• [08/2023] 🔥 We introduced InsTag, our SFT data analysis tool. Check out the paper.

• Model Checkpoints
• Citation

InsTagger is a LLaMa-2 based SFT model trained with FastChat in the vicuna template. You can easily download weight at HuggingFace ModelHub and then use FastChat to serve or inference. Demo codes are about to be released.

• InsTagger for local query tagging:

  InsTagger is an tagging LLM which is fine-tuned on InsTag's tagging results on open-resourced SFT data. The model is based on 7B version LLaMA-2.

  Download the model checkpoint below:

  Model	Checkpoint	Exact Match F1	Semantic-based Fuzzy Match F1	License
  LocalTagger	🤗 HF Link	31.8%	73.4%	LLaMA 2 License

• TagLM, fine-tuned on our SFT data sub-sampled by complexity-first diverse sampling procedure:

  With only 6k data from current open-resourced SFT dataset, TagLM can outperform many open-resourced LLMs on MT-Bench using GPT-4 as a judge.

  Download the model checkpoint below:

  Model	Checkpoint	MT-Bench	License
  TagLM-13B-v1.0	🤗 HF Hub Link	6.44	LLaMA License
  TagLM-13B-v2.0	🤗 HF Hub Link	6.55	LLaMA 2 License

  All models are either based on LLaMA or LLaMA-2 and should be used under their licenses accordingly. All the models are fine-tuned using FastChat codebase, and we apply the system template of Vicuna V1.1.

Please cite our work if you find the repository helpful.

@misc{lu2023instag,
      title={#InsTag: Instruction Tagging for Analyzing Supervised Fine-tuning of Large Language Models}, 
      author={Keming Lu and Hongyi Yuan and Zheng Yuan and Runji Lin and Junyang Lin and Chuanqi Tan and Chang Zhou and Jingren Zhou},
      year={2023},
      eprint={2308.07074},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}