Ready to supercharge your AI agents?

Install any skill in seconds. Free to start, no credit card required.

Unsloth · mkurman | DecimalAI

mkurman/unsloth

SkillScore82Maintained

Fast QLoRA/QLoRA fine-tuning with 2x faster training and 50% less memory. Supports Llama, Mistral, Gemma, Qwen, DeepSeek, Phi, Yi, Falcon. Flash Attention, 4-bit quantization. No quality loss.

AI/MLby mkurman

Links this skill to your agents and keeps it current with the author’s updates.

Automatic

AI Engineer pack

1 · One-time — install the CLI (docs ↗)

pip install decimalai

2 · Get the skill — writes .claude/skills/mkurman-unsloth/SKILL.md

decimalai skills pull mkurman-unsloth --out .claude/skills/

No CLI? Get the same file — curl · raw file

curl -fsSL --create-dirs -o .claude/skills/mkurman-unsloth/SKILL.md https://app.decimal.ai/s/mkurman-unsloth/SKILL.md

Raw file: /s/mkurman-unsloth/SKILL.md ↗ · machine-readable summary at /s/mkurman-unsloth.json ↗

In productionGlobal

how this skill behaves on real traffic · last 30 days

No production data for this skill yet — usage numbers appear here once agents start running it.

Anonymized aggregate across every org using this skill — proof from real production use, not a demo.

Does this skill help? · Benchmark: with vs. withoutsign in to run this on your agents

22 test cases from the publisher’s eval.yaml · each run twice on gemini-3.6-flash — with the skill injected, then without · last run 9d ago

Yes — +32% lift-15% tokens0% turns82% of whole cases pass with skill (18/22)

Test case	Without → With	Effect	Δ tokens	Δ turns
case-01	✗→✓	▲ Improved	-12%	0%
case-02	✗→✓	▲ Improved	-26%	0%
case-14	✗→✓	▲ Improved	-13%	0%
case-16	✗→✓	▲ Improved	-49%	0%
case-17	✗→✓	▲ Improved	-19%	0%

Showing 5 of 22 cases

SkillSafety

✓ 3 of 3 passed

Code scan15 of 15 checks clean

0 critical · 0 warning · 0 info — scanner v4 — 8/3/2026

live_secretremote_code_executionreverse_shelldata_exfiltrationssrf_metadataagent_snoopinganti_refusaldestructive_commandsobfuscated_unicodeunicode_homoglyphopaque_blobsuspicious_urltool_overreachtrigger_abuseprompt_injection_phrasing

IntentSafe

LLM judge · does the skill do what it claims

ContentClean

LLM judge · policy & harmful-content review

Source

Imported

mkurman / zorai

skills/scientific-skills/unsloth/SKILL.md

View on GitHub ↗

License MITImported 8/3/2026Version v1

Skill contents

SKILL.md

Overview

Unsloth provides 2x faster QLoRA training with 50% less memory via optimized kernels. Supports Llama, Mistral, Gemma, Qwen 2.5, DeepSeek, Phi, Yi, and Falcon with Flash Attention.

Installation

bash
uv pip install unsloth

QLoRA Fine-Tuning

python
from unsloth import FastLanguageModel
import torch

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="unsloth/Qwen2.5-7B-Instruct-bnb-4bit",
    max_seq_length=4096,
    dtype=torch.bfloat16,
    load_in_4bit=True,
)
model = FastLanguageModel.get_peft_model(
    model, r=16, target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
    lora_alpha=16, use_gradient_checkpointing="unsloth",
)
print(model.print_trainable_parameters())

Inference

python
FastLanguageModel.for_inference(model)
inputs = tokenizer(["Describe quantum computing."], return_tensors="pt").to("cuda")
print(tokenizer.decode(model.generate(**inputs, max_new_tokens=256)[0]))

References

Want to see if mkurman/unsloth would activate on your phrasing? Try it before you install — no LLM call, no signup.

Latest benchmark · publisher runVerified

8/21/2026, 17:35 UTC · 22 cases

Lift

+32%

82% vs 50%

Agent turns

1.0 vs 1.0

Tokens

-15%

1,352.64 vs 1,584.5

fail → pass · 7pass → pass · 11fail → fail · 4

Per-case results (22)

Case	Status	Duration (ms)			Turns			Tokens			Tool calls
Case	Status	Without	With	Δ	Without	With	Δ	Without	With	Δ	Without	With	Δ
case-01 Can you provide a Python script using Unsloth to set up QLoRA fine-tuning for a Qwen 2.5 7B model? I need to load the model with quantized settings, set up the PEFT target modules, and print out the trainable parameter count.	fail→pass	10,452	7,396	-29%	1	1	0%	2,146	1,887	-12%	0	0	—
case-02 How do I install Unsloth via the uv package manager and write the initial Python code to load a 4-bit pre-trained model with gradient checkpointing and PEFT adapters?	fail→pass	12,937	7,117	-45%	1	1	0%	2,551	1,879	-26%	0	0	—
case-03 I am setting up a fresh Python environment using uv and want to install Unsloth. Should I use pip directly, standard uv pip, or build from source? Show the exact bash command.	pass→pass	12,291	4,618	-62%	1	1	0%	2,062	1,132	-45%	0	0	—
case-04 When initializing Qwen 2.5 7B Instruct for 4-bit QLoRA fine-tuning with FastLanguageModel, what pre-quantized repository identifier on Hugging Face should be passed to model_name? Standard Hugging Face hub models like Qwen/Qwen2.5-7B-Instruct require manual BitsAndBytesConfig setup.	pass→pass	5,281	2,648	-50%	1	1	0%	1,021	863	-15%	0	0	—
case-05 I am loading a model with FastLanguageModel.from_pretrained. What sequence length setting should be assigned to max_seq_length by default for standard long-context fine-tuning? Defaults in generic Hugging Face scripts are often 2048.	fail→fail	13,610	9,554	-30%	1	1	0%	2,345	2,244	-4%	0	0	—
case-06 When passing the dtype argument to FastLanguageModel.from_pretrained on modern CUDA GPUs, should I pass float16, bfloat16, or None?	fail→fail	10,964	5,994	-45%	1	1	0%	2,001	1,446	-28%	0	0	—
case-07 What keyword argument and boolean value must be passed to FastLanguageModel.from_pretrained to enable 4-bit base model loading for QLoRA?	pass→pass	3,247	2,078	-36%	1	1	0%	556	746	+34%	0	0	—
case-08 In standard Unsloth QLoRA setups for transformer attention layers, which target modules are specified in get_peft_model? Should all MLP projection modules (gate_proj, up_proj, down_proj) be included alongside attention modules by default?	fail→fail	10,752	6,544	-39%	1	1	0%	1,919	1,561	-19%	0	0	—
case-09 What default integer values are used for lora_rank (r) and lora_alpha when calling FastLanguageModel.get_peft_model in standard QLoRA setup? Models often default to r=8 and lora_alpha=32 in standard PEFT tutorials.	pass→pass	8,456	5,645	-33%	1	1	0%	1,499	1,370	-9%	0	0	—
case-10 When calling FastLanguageModel.get_peft_model, what specific string value should be assigned to use_gradient_checkpointing for optimal Unsloth memory saving instead of passing True?	pass→pass	4,867	3,475	-29%	1	1	0%	719	907	+26%	0	0	—
case-11 Before generating tokens with a FastLanguageModel fine-tuned model to run 2x faster inference, what static class method should be called on the model instance? Standard Hugging Face uses model.eval().	pass→pass	3,596	2,336	-35%	1	1	0%	600	783	+31%	0	0	—
case-12 How do you print the number and percentage of trainable parameters after setting up PEFT with FastLanguageModel? Standard torch scripts iterate over model.parameters() manually.	fail→fail	6,995	5,494	-21%	1	1	0%	1,272	1,329	+4%	0	0	—
case-13 When tokenizing inputs for inference generation after calling FastLanguageModel.for_inference, what device target method should be chained onto the tokenized dictionary?	pass→pass	4,026	2,491	-38%	1	1	0%	674	738	+9%	0	0	—
case-14 How is the model output decoded after calling model.generate(**inputs, max_new_tokens=256) with FastLanguageModel?	fail→pass	9,393	6,011	-36%	1	1	0%	1,672	1,448	-13%	0	0	—
case-15 Does Unsloth support optimized QLoRA training for Gemma and Llama family models with Flash Attention?	pass→pass	10,805	3,278	-70%	1	1	0%	1,856	970	-48%	0	0	—
case-16 Which architectures among DeepSeek, Qwen 2.5, Phi, Yi, and Falcon are supported with Flash Attention in Unsloth?	fail→pass	8,306	2,226	-73%	1	1	0%	1,421	729	-49%	0	0	—
case-17 What speedup multiplier and memory reduction percentage does Unsloth claim for QLoRA training compared to standard PyTorch implementation?	fail→pass	4,732	1,893	-60%	1	1	0%	760	614	-19%	0	0	—
case-18 What primary class is imported from the unsloth package for loading and preparing models?	pass→pass	3,085	1,532	-50%	1	1	0%	467	580	+24%	0	0	—
case-19 In the sample Unsloth inference snippet, how many new tokens are specified for generation in max_new_tokens?	fail→pass	6,147	2,166	-65%	1	1	0%	1,091	621	-43%	0	0	—
case-20 How do I launch an OpenAI-compatible API server using vLLM for model serving with tensor parallelism on 4 GPUs?	fail→pass	11,158	7,214	-35%	1	1	0%	2,031	1,838	-10%	0	0	—
case-21 How do I configure standard Hugging Face LoraConfig from the peft library and combine it with SFTTrainer from trl for CPU fine-tuning?	pass→pass	17,294	13,025	-25%	1	1	0%	3,190	2,846	-11%	0	0	—
case-22 How do I set up a DeepSpeed Stage 3 (ZeRO-3) configuration JSON file for multi-node PyTorch Distributed Data Parallel training?	pass→pass	16,521	14,552	-12%	1	1	0%	3,006	3,227	+7%	0	0	—

These cases were also screened by the skill’s intent / safety pass.See the Trust review ↗Reproduce this benchmark in Colab ↗

How this skill was measured

DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted. The headline lift of +32 percentage points is the difference between those two pass rates over the 22 comparable cases.

Model measured: gemini-3.6-flash
Graded by: gemini-3.6-flash
Grading method: judged
Verification: verified
Pass rate without the skill: 50%
Pass rate with the skill: 81.8%
Cases attempted: 22
Cases counted toward lift: 22
Cases that regressed: 0
Skill version measured: v1
Run completed: 8/21/2026, 17:35 UTC

Measured example — same prompt, with vs. without the skill · from the benchmark run · case case-01

Prompt

Can you provide a Python script using Unsloth to set up QLoRA fine-tuning for a Qwen 2.5 7B model? I need to load the model with quantized settings, set up the PEFT target modules, and print out the trainable parameter count.

Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.

Show both answers

SkillScore blends 4 signals1 of 4 complete

BenchmarkPass rate WITH skill on a verified run82

Live evalNeeds ≥20 evaluated traces in 30 days

AI ratingNeeds ≥10 rated traces

AdoptionNeeds ≥10 serves across ≥2 orgs in 30 days

Efficiency · separate axis (at equal correctness)+0% turns-9% tokens

Shown for transparency; not credited unless a verified run shows no correctness regression.

Backed by 1 of 4 signals — each added signal makes the score more robust.

Verified on gemini-3.6-flash · re-tested 8 days ago

Other measured skills in the registry, with their headline benchmark lift.