Research map / Large language models / Instruction and fine-tuning
#InsTag: Instruction Tagging for Analyzing Supervised Fine-tuning of Large Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes InsTag, an open-set ChatGPT-based instruction tagger with tag normalization, uses its tags to analyze SFT dataset diversity and complexity, and selects 6K samples to fine-tune LLaMA/LLaMA-2 13B models (TagLM) that reach MT-Bench scores of 6.44 and 6.55.
Problem
Diversity and complexity are considered key to good SFT data for instruction following, but they lack quantitative definitions and analyses.
Approach
Use open-set fine-grained intention tags produced by ChatGPT, normalized via frequency filtering, rule/semantic/association aggregation, to quantify SFT query diversity (tag coverage) and complexity (average tags per query), and select a complexity-first diverse 6K subset for fine-tuning.
Key results
- TagLM-13b-v1.0 (LLaMA-13B, 6K samples) achieves MT-Bench 6.44, outperforming open models trained on far more data such as Vicuna-13b-v1.3 (125K, 6.39).
- TagLM-13b-v2.0 (LLaMA-2-13B, 6K samples) achieves MT-Bench 6.55, only 0.1 below Llama-2-13b-chat (6.65, which uses RLHF).
- InsTag tag quality: GPT-4 judged 96.1% tag precision and 86.6% tag consistency; human majority vote 100% on 1% sample.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fine-tune LLaMA-13B on InsTag-selected, diverse and complex 6K samples rather than larger open-source SFT sets | LLaMA-13B · MT-Bench | score: 6.39 -> 6.44 | Helped |
| Increase average tag number in the 6K training subset while keeping tag coverage at 100% (ablation) | LLaMA-13B fine-tuned on pooled SFT data · MT-Bench | score | Helped |
| Increase tag coverage in the 6K training subset while keeping average tag number around 5.0 (ablation) | LLaMA-13B fine-tuned on pooled SFT data · MT-Bench | score | Helped |
Surprising findings
- Human-human Fleiss kappa for tag precision was only 0.47, indicating low agreement among annotators despite high majority-vote accuracy.
- On fine-grained levels, the complexity-performance trend becomes less evident, and math and code datasets show high complexity but low diversity.
Limitations
- MT-Bench is the main evaluation and GPT-4 judgments are noisy, so results may miss SFT data influences.
- Analysis of SFT datasets focuses on English only.
- Fine-grained tag-count complexity signal becomes less reliable at fine granularity since ChatGPT may miss or filter tags.
Experimental setup
Datasets: ShareGPT OpenChat UltraChat Alpaca WizardLM FLAN Dolly OAssist Unnatural Lima GSM8K MATH
Benchmarks: MT-Bench AlpacaEval
Metrics: MT-Bench score (GPT-4 judged) tag precision tag consistency tag-level F1 (exact and fuzzy match)
Baselines: GPT-4 GPT-3.5-turbo Claude-v1 Llama-2-13b-chat Vicuna-13b-v1.3 Vicuna-13b-v1.1 WizardLM-13b Baize-v2-13b OpenChat-13b-v1 Alpaca-13b
Builds on: Vicuna Evol-Instruct Self-Instruct PhraseBERT FP-Growth
Closest papers by meaning
- MAmmoTH2: Scaling Instructions from the Web (NeurIPS 2024 · similarity 0.90)
- From Instance Training to Instruction Learning: Task Adapters Generation from Instructions (NeurIPS 2024 · similarity 0.89)
- Chain-of-Instructions: Compositional Instruction Tuning on Large Language Models (AAAI 2025 · similarity 0.89)
- LESS: Selecting Influential Data for Targeted Instruction Tuning (ICML 2024 · similarity 0.89)
- The Best Instruction-Tuning Data are Those That Fit (NeurIPS 2025 · similarity 0.89)
- A Closer Look at the Limitations of Instruction Tuning (ICML 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.