Model tools
Free, open-source projects for changing a model after you download it: remove refusals, fine-tune it on your own data, shrink it to fit your hardware, or merge it with another model.
Before you run anything: these are third-party projects listed from public GitHub data, not tested by us. They run code on your computer. Read the code, license and instructions first, and try them in an isolated environment.
GitHub details checked: 2026-10-21 · 23:41:51 UTC.
These tools edit a model's weights to suppress refusal behavior, usually without retraining. Results vary by architecture. Removing refusals doesn't make a model more accurate.
p-e-w/heretic
Removes refusal behavior from transformer models automatically. It pairs directional ablation (abliteration) with an Optuna parameter search that tries to cut refusals while keeping the model close to the original, so you don't have to pick layers by hand.
★ 36,109 · Last updated 2026-10-20 · License: AGPL-3.0
Summary source: README ↗
Linked from these model cards:
elder-plinius/OBLITERATUS
A larger refusal-removal toolkit with several abliteration methods, usable locally or through a hosted Hugging Face Space. Check the telemetry setting before you run it: the README describes telemetry as opt-in, but on by default in the Spaces version.
★ 8,902 · Last updated 2026-10-21 · License: AGPL-3.0
Summary source: README ↗
Sumandora/remove-refusals-with-transformers
A small proof-of-concept that computes a refusal direction and removes it using plain Hugging Face Transformers instead of TransformerLens. The author says it was mostly tested on models under 3B and that some custom architectures, including some Qwen implementations, may not work.
★ 2,299 · Last updated 2025-11-27 · License: Apache-2.0
Summary source: README ↗
Linked from these model cards:
FailSpy/abliterator
A barebones Python library for abliteration experiments in TransformerLens: caching activations, calculating refusal directions and testing orthogonalized weights. The author describes its documentation as slim.
★ 773 · Last updated 2024-06-11 · License: MIT
Summary source: README ↗
Train a model further on your own data or preferences. LoRA and QLoRA train a small set of extra weights, so they need far less memory than full fine-tuning.
unslothai/unsloth
A desktop app and Python library for running and training models locally. Its README lists LoRA, QLoRA, full fine-tuning, DPO and GRPO, plus export to formats including GGUF, on Windows, Linux, WSL and macOS.
★ 78,375 · Last updated 2026-10-21 · License: Apache-2.0
Summary source: README ↗
hiyouga/LlamaFactory
Fine-tunes more than 100 language and vision-language models from a command line or a browser-based Web UI, aimed at training without writing code.
★ 75,635 · Last updated 2026-10-21 · License: Apache-2.0
Summary source: README ↗
huggingface/peft
Hugging Face's library of parameter-efficient fine-tuning methods such as LoRA. They train a small number of extra parameters instead of the whole model, which cuts compute and storage costs.
★ 21,822 · Last updated 2026-10-21 · License: Apache-2.0
Summary source: README ↗
huggingface/trl
Hugging Face's library for post-training models with supervised fine-tuning (SFT), DPO, GRPO and related methods, built on Transformers.
★ 19,565 · Last updated 2026-10-21 · License: Apache-2.0
Summary source: README ↗
Linked from these model cards:
axolotl-ai-cloud/axolotl
An open-source fine-tuning framework that runs preprocessing, training, evaluation and quantization from one YAML config file. It supports LoRA and QLoRA, and recent releases added GGUF export for llama.cpp and Ollama.
★ 12,606 · Last updated 2026-10-21 · License: Apache-2.0
Summary source: README ↗
Linked from these model cards:
ml-explore/mlx-lm
Apple's MLX package for running and fine-tuning models on Apple silicon. It can quantize models and upload them to the Hugging Face Hub, and supports LoRA and full fine-tuning.
★ 7,431 · Last updated 2026-10-20 · License: MIT
Summary source: README ↗
Goekdeniz-Guelmez/mlx-lm-lora
Trains models locally on Apple Silicon with MLX: LoRA, DoRA, QLoRA or full fine-tuning, with SFT, DPO and other preference methods. It works with models MLX-LM supports, including Llama, Mistral, Qwen and Gemma.
★ 437 · Last updated 2026-10-13 · License: Apache-2.0
Summary source: README ↗
Shrink a model's weights to lower precision so it uses less memory, or convert it to the format your runtime expects. Lower precision can cost some accuracy.
ggml-org/llama.cpp
Runs models locally and includes the tools behind most GGUF files. A conversion script turns Hugging Face weights into GGUF, and llama-quantize shrinks a high-precision GGUF to lower-bit formats, which can cost some accuracy.
★ 131,943 · Last updated 2026-10-21 · License: MIT
Summary source: README ↗ · tools/quantize/README.md ↗
ml-explore/mlx-lm
Apple's MLX package for running and fine-tuning models on Apple silicon. It can quantize models and upload them to the Hugging Face Hub, and supports LoRA and full fine-tuning.
★ 7,431 · Last updated 2026-10-20 · License: MIT
Summary source: README ↗
vllm-project/llm-compressor
Quantizes and compresses Hugging Face models for serving with vLLM, saving results in the compressed-tensors format.
★ 3,897 · Last updated 2026-10-21 · License: Apache-2.0
Summary source: README ↗
turboderp-org/exllamav3
An inference library for local GPUs with its own EXL3 quantization format and CPU offloading for large mixture-of-experts models. The project recommends TabbyAPI as its server.
★ 1,723 · Last updated 2026-10-20 · License: MIT
Summary source: README ↗
ModelCloud/GPTQModel
A quantization toolkit covering GPTQ, AWQ, GGUF, FP8, EXL3 and other formats on NVIDIA, AMD, Intel and Apple hardware.
★ 1,274 · Last updated 2026-10-21 · License: not stated
Summary source: README ↗
Combine the weights of compatible models into a new model without full retraining.
arcee-ai/mergekit
Merges pretrained models with many merge methods. It works out-of-core, so merges can run on a CPU or with as little as 8 GB of VRAM, and it also handles LoRA extraction and mixture-of-experts merges.
★ 7,411 · Last updated 2026-10-20 · License: LGPL-3.0
Summary source: README ↗
Free material that explains these techniques step by step.
mlabonne/llm-course
A free course with roadmaps and Colab notebooks. Its notebooks include one-click tools for merging with MergeKit, quantizing to GGUF, GPTQ, EXL2, AWQ and HQQ, and automatic abliteration with custom datasets.
★ 84,565 · Last updated 2026-02-05 · License: Apache-2.0
Summary source: README ↗