Research map / Large language models / Instruction and fine-tuning

SWIFT: A Scalable Lightweight Infrastructure for Fine-Tuning

AAAI 2025 · Large language models › Instruction and fine-tuning · approach: Fine-tuning generalization and delta compression · 40 citations (OpenAlex)

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.

Presents SWIFT, an open-source one-stop ModelScope framework for pre-training, fine-tuning, alignment, RFT, quantization, inference, and evaluation of LLMs and MLLMs, with agent fine-tuning experiments on ToolBench.

Problem

Existing LLM/MLLM training frameworks support only a limited set of models and techniques, lacking support for new models and end-to-end post-training workflows, making fine-tuning hard for developers.

Approach

A unified, extensible training infrastructure that standardizes data formats, model templates, and tuners across pure-text and multi-modal LLMs, and integrates post-training steps (quantization, export, inference, deployment, evaluation) in one framework, with loss-scale weighting for agent fine-tuning.

Key results

• Qwen2-7B-Instruct fully fine-tuned on agent data reaches Act.EM 60.01 in-domain versus 54.74 for the original model, with hallucination rate reduced from 4.16 to 2.58.
• LLaMA3-8B-Instruct with LoRA and loss-scale improves in-domain Act.EM from 36.17 to 58.15 and reduces hallucination rate from 15.68 to 1.57 in the ablation.
• In the tuner benchmark on Qwen-7B-Chat, LISA gives the lowest memory (31.11 GiB) and highest speed (2.66 samples/s) versus full training at 73.53 GiB.

Measured effects reported in the paper

Surprising findings

• The LLaMA3-8B LoRA agent model's hallucination rate in the main table (15.11%) is higher than the original model's rate in-domain, contrasting with the reported reduction claims for other models.
• Qwen2-7B LoRA agent training yields the lowest hallucination rate (0.9%) in-domain, lower than GPT-4's 5.98%.

Limitations

• Experiments are limited to agent fine-tuning on ToolBench with a few 7-8B models; no broad benchmark validation of the framework itself.
• Megatron support does not fully cover mainstream LLMs and MLLMs, as acknowledged by the authors.
• Many claims on model/dataset counts are not supported by detailed evaluation, and several tables are inconsistent (e.g., 300+ vs 550+ LLMs).

Experimental setup

Datasets: ToolBench AgentFlan MSAgent-Pro MS-Agent alpaca-en firefly-train-1.1M llava-instruct-150k refcoco/refcocog

Benchmarks: ToolBench (in-domain) ToolBench (out-of-domain)

Metrics: Plan.EM Act.EM Hallucination rate Avg.F1 ROUGE-L training loss memory (GiB) speed (samples/s)

Baselines: LLaMA-Factory FireFly FastChat Axolotl LMFlow GPT-4

Builds on: PEFT TRL Transformers Megatron vLLM

Closest papers by meaning

• A Paradigm Shift in Machine Translation: Boosting Translation Performance of Large Language Models (ICLR 2024 · similarity 0.90)
• JudgeLM: Fine-tuned Large Language Models are Scalable Judges (ICLR 2025 · similarity 0.90)
• When Scaling Meets LLM Finetuning: The Effect of Data, Model and Finetuning Method (ICLR 2024 · similarity 0.89)
• Data Efficient Adaptation in Large Language Models via Continuous Low-Rank Fine-Tuning (NeurIPS 2025 · similarity 0.89)
• Unveiling the Secret Recipe: A Guide For Supervised Fine-Tuning Small LLMs (ICLR 2025 · similarity 0.89)
• LongLoRA: Efficient Fine-tuning of Long-Context Large Language Models (ICLR 2024 · similarity 0.89)