Firefly: Pretrain and fine-tune large language models

Understand the project

Practical overview

Use Firefly when you need a configurable toolkit for model training or fine-tuning workflows, especially instruction tuning with LoRA or QLoRA.

Pretrain a supported large language model.
Fine-tune a model on instruction data using full-parameter, LoRA or QLoRA methods.
Run direct preference optimization (DPO).
Use project-provided instruction datasets and released fine-tuned model weights as related resources.

Before you choose

The project is published as an open-source public GitHub repository. Its technical profile identifies Linux and server environments.

The README lists support for model families including Qwen, Yi, Llama, Gemma, MiniCPM, DeepSeek, Phi, InternLM, Baichuan, ChatGLM, Mistral and Mixtral; verify the repository configuration for the exact model variant.
Training methods include full-parameter updates, LoRA and QLoRA; QLoRA is recommended by the project for instruction fine-tuning with limited resources.
The repository describes configuration-file-based training and model-specific chat-template adaptation for instruction fine-tuning and DPO.
Unsloth support is described as an option for accelerating training and reducing memory use; reported gains are project results, not a guarantee for an individual setup.
The README links multiple instruction datasets and Firefly fine-tuned weights as separate project resources.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Choose a supported model and training approach through Firefly’s configuration-file workflow.

A model supported by Firefly, such as Qwen, Llama, Gemma, MiniCPM, Yi, or DeepSeek.
A training task: pretraining, supervised fine-tuning (SFT), or DPO.
A training method: full-parameter training, LoRA, or QLoRA.

Select a supported model family for the intended training task.
Choose pretraining, SFT, or DPO as the training objective.
Choose full-parameter training, LoRA, or QLoRA; Firefly recommends QLoRA for instruction fine-tuning when training resources are limited.
Use Firefly’s configuration-file workflow to configure training for the selected model.

A Firefly configuration-based training workflow is specified for the chosen model, objective, and training method.

The README does not provide the configuration schema, specific file names, setup instructions, or commands in the documented material.
Supported models and training options vary; verify the desired model against the project documentation.
The resource requirements for a particular configuration are not specified.

Evidence and transparency

Source record details

Source availability: Active
Sources checked: 15 October 2026
Original website title: yangjianxin1/Firefly · Official homepage title · retained in its source language