Verified project record

yangjianxin1/Firefly

Firefly provides configurable workflows for pretraining, instruction fine-tuning, and DPO across many mainstream open large language models. It supports full-parameter, LoRA, and QLoRA training modes to accommodate varying resource requirements.

Project overview

Firefly stands out for combining full-parameter, LoRA, and QLoRA workflows in a single configuration-driven codebase, enabling practitioners to adapt mainstream open LLMs without building separate training pipelines.

Project type: Model Development · Model Runtime
Deployment: Refer to project documentation
License: License pending

Best for

AI engineers and developers who need configurable workflows for pretraining, instruction fine-tuning, or DPO across mainstream open LLMs.
Practitioners who need full-parameter, LoRA, or QLoRA training modes in a single project.

Key capabilities

Runs pretraining, instruction fine-tuning, and DPO workflows for supported mainstream open LLMs.
Offers full-parameter training plus LoRA and QLoRA efficient-training modes.
Uses configuration files for training different models, allowing users to inspect, modify, or add model training configurations.
Provides a multi-turn chat script compatible with models trained by the project.
Accepts JSONL training data with a required conversation field and documents reusable instruction datasets in a unified format.
Supports 4-bit inference to reduce VRAM requirements, with a documented quality trade-off.

Limitations and risks

The project's released models may produce incorrect responses for factual knowledge.
The project's released models have remaining shortcomings in coding and reasoning ability.
4-bit inference lowers VRAM requirements but may slightly reduce quality.
Before adoption, verify that downstream safeguards are in place. The project's released models may generate discriminatory, harmful, or unethical statements because they have not undergone harmlessness fine-tuning.

Getting started

Setup is rated hard. Model-specific dependency versions and compatibility exceptions are documented. Training requires selecting or modifying a configuration and running Python, DeepSpeed, or torchrun commands.

Evidence and sources

README: **Firefly** 是一个开源的大模型训练项目，支持对主流的大模型进行预训练、指令微调和DPO，包括但不限于Qwen2、Yi-1.5、Llama3、Gemma、Qwen1.5、MiniCPM、MiniCPM3、Llama、InternLM、Baichuan、ChatGLM、Yi、Deepseek、Qwen、Orion、Ziya、Xverse、Mistr…
README: 本项目支持**全量参数训练、LoRA、QLoRA高效训练**，支持**预训练、SFT、DPO**。如果你的训练资源有限，我们极力推荐使用QLoRA进行指令微调，因为我们在Open LLM Leaderboard上验证了该方法的有效性，并且取得了非常不错的成绩。
README: 训练时的参数配置存储在train_args目录下，方便统一管理和更改。大家可以在train_args目录下查看不同模型的训练配置，按需修改或添加。
README: 我们提供了多轮对话的交互脚本，详见script/chat目录，该脚本可同时兼容本项目训练的所有模型进行推理。
README: 训练数据为jsonl格式，每一行的数据格式如下，其中conversation字段是必需的，可以根据实际需求添加或删除其他字段。

Reviewed at: 2026-07-16T13:11:34Z