Choose theme
AI Resources
LLaMA Factory
LLaMA Factory is a framework for fine-tuning language and vision-language models through configuration files and a web UI.
Its examples cover training, inference and export. Those are separate steps: choosing a dataset, producing a LoRA adapter and loading a complete model are different parts of the workflow. Use this as a first read, not a recommendation. Open the original project before trusting details like terms, limits, privacy, cost, setup, or safety.
What it is
Configurable model training
A CLI and Gradio interface for fine-tuning supported models, with examples for LoRA, quantized LoRA and full-parameter training.
Why it stands out
An adapter has a matching base
The LoRA example's inference configuration names both the original model and the saved adapter, while keeping its chat template.
Availability
Code and runnable examples
The repository provides dataset-format instructions and separate training, chat and export configurations. They still require a model, data and a suitable compute environment.
Why it matters
What makes it useful
To train on your own instruction-and-answer examples, make the dataset visible to the loader. The data guide asks readers to register it in dataset_info.json and select that registered name in the training configuration. Its column mappings connect your field names to prompts and responses. Dropping a JSON file into a folder alone does not describe how the trainer should read it.
What to know
Where it fits
For a LoRA result, keep track of the base model as well as the adapter directory. The Qwen3 example trains into saves/qwen3-4b/lora/sft; its chat configuration loads that adapter alongside Qwen/Qwen3-4B-Instruct-2507 and the same qwen3_nothink template. That directory is not presented as a self-contained replacement for the full base model.
Notable points
What stands out
The repository's estimated memory table gives different 7B examples: 60 GB for full training with pure_bf16, 16 GB for 16-bit LoRA, and 6 GB for 4-bit QLoRA. The ordinary bf16/fp16 full-training row is 120 GB. Read the method and precision with the number; the table is an estimate, not a promise that a particular GPU setup will fit.
Before using
What to review
The published stable release is v0.9.5; a newer development checkout is a different version choice.
For older installations, the project's Chat API and reward-modeling WebUI advisories list versions through 0.9.3 as affected and 0.9.4 as patched. The first concerns supplied media URLs and local files; the second concerns loading a value-head checkpoint. Compare the installed version and the workflow with those published ranges rather than treating a web UI or local process as a safety guarantee.
Reader fit
Who may find it relevant
If you need to assess a trained assistant on held-out examples, distinguish training progress from evaluation. The sample Qwen3 LoRA file enables a loss plot, but its eval_dataset, val_size and evaluation schedule are commented out. That sample does not enable those checks just by being a fine-tuning recipe. Choose an evaluation configuration before treating a training curve as evidence of the result's usefulness.
Editorial note
Why LifeHubber lists it
When exporting a merged model, separate the merge configuration from the quantized training configuration. The examples guide and merge recipe warn against using a quantized base or quantization_bit while merging LoRA adapters. Their merge example loads the base model with the saved adapter. A low-memory QLoRA training choice therefore does not carry unchanged into the documented merge step.
Source links
Source materials
Reader note
Before relying on this entry
LifeHubber lists entries to help readers inspect AI projects, not to endorse them or prove they are safe, suitable, accurate, maintained, or right for a specific use. We do not verify every entry in depth. Before relying on anything listed, review the original materials, terms, privacy practices, limits, and risks that matter for your situation.
More in Ecosystem
Keep browsing this category
Explore more AI ecosystem resources.
Laya
NandhaKishorM/laya
An early Apache-2.0 model family and Python runtime for bounded choice, score, and yes-or-no decisions, with English, multilingual, and task-specialized checkpoints plus a router that selects between them.
AnyJev
nokia-applied-research/AnyJev
An Apache-2.0 Python framework that turns existing Transformers or vLLM-served language models into bounded choice, yes-or-no, and score decisions, with levels for zero-label bias correction, calibration, and question-specific decision heads.
CubeSandbox
TencentCloud/CubeSandbox
TencentCloud sandbox infrastructure for AI agents, with KVM MicroVM execution, E2B-compatible code workflows, templates, snapshots, an E2B-compatible Volume framework, preview Kubernetes deployment, and preview S3-backed cross-node pause and resume.
For project maintainers
Listed here? You can use the badge.
If you maintain a project with a current LifeHubber listing, you may add the optional “Listed on LifeHubber AI Resources” badge to its README, docs, or website. No introduction or permission request is needed.
