Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Streamlined fine-tuning framework for LLMs. Supports full fine-tune, LoRA, QLoRA, FSDP, DeepSpeed, and multi-GPU. YAML config driven. Works with Llama, Mistral, Qwen, DeepSeek, and hundreds of HF models.
.claude/skills/mkurman-axolotl/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-17 | ✗→✓ | ▲ Improved | -54% | 0% |
| case-02 | ✗→✓ | ▲ Improved | -51% | 0% |
| case-04 | ✗→✓ | ▲ Improved | -8% | 0% |
| case-11 | ✗→✓ | ▲ Improved | -55% | 0% |
| case-12 | ✗→✓ | ▲ Improved | -49% | 0% |
Axolotl is a fine-tuning framework supporting SFT, QLoRA, LoRA, full fine-tuning, DPO, and multimodal tuning for 100+ models (Llama, Mistral, Qwen, Gemma, DeepSeek). YAML-driven config avoids boilerplate. Supports multi-GPU, FSDP, DeepSpeed, and flash attention.
bashgit clone https://github.com/OpenAccess-AI-Collective/axolotl cd axolotl uv pip install -e .
yaml# config.yml base_model: Qwen/Qwen2.5-1.5B-Instruct model_type: AutoModelForCausalLM tokenizer_type: AutoTokenizer output_dir: ./output # LoRA adapter: lora lora_r: 16 lora_alpha: 32 lora_dropout: 0.05 lora_target_modules: - q_proj - v_proj # Training sequence_len: 2048 micro_batch_size: 2 gradient_accumulation_steps: 4 num_epochs: 3 learning_rate: 2e-5 optimizer: adamw_bnb_8bit
bashaccelerate launch -m axolotl.cli.train config.yml
bashpython -m axolotl.cli.inference --lora_model_dir ./output --base_model Qwen/Qwen2.5-1.5B-Instruct
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-17 | fail→pass | 8,062 | 1,604 | -80% | 1 | 1 | 0% | 1,445 | 663 | -54% | 0 | 0 | — |
case-01 | fail→fail | 12,803 | 9,943 | -22% | 1 | 1 | 0% | 2,543 | 2,375 | -7% | 0 | 0 | — |
case-02 | fail→pass | 8,886 | 2,164 | -76% | 1 | 1 | 0% | 1,638 | 796 | -51% | 0 | 0 | — |
case-16 | pass→pass | 5,369 | 2,159 | -60% | 1 | 1 | 0% | 826 | 789 | -4% | 0 | 0 | — |
case-03 | fail→fail | 8,901 | 5,215 | -41% | 1 | 1 | 0% | 1,672 | 1,359 | -19% | 0 | 0 | — |
case-04 | fail→pass | 11,552 | 7,895 | -32% | 1 | 1 | 0% | 2,236 | 2,061 | -8% | 0 | 0 | — |
case-05 | pass→pass | 9,589 | 6,453 | -33% | 1 | 1 | 0% | 1,876 | 1,751 | -7% | 0 | 0 | — |
case-06 | pass→pass | 10,353 | 7,035 | -32% | 1 | 1 | 0% | 2,064 | 1,753 | -15% | 0 | 0 | — |
case-07 | fail→fail | 2,473 | 1,470 | -41% | 1 | 1 | 0% | 406 | 609 | +50% | 0 | 0 | — |
case-08 | pass→pass | 3,492 | 2,327 | -33% | 1 | 1 | 0% | 562 | 783 | +39% | 0 | 0 | — |
case-09 | pass→pass | 5,130 | 3,359 | -35% | 1 | 1 | 0% | 925 | 913 | -1% | 0 | 0 | — |
case-10 | pass→pass | 3,004 | 1,973 | -34% | 1 | 1 | 0% | 509 | 658 | +29% | 0 | 0 | — |
case-11 | fail→pass | 8,598 | 2,405 | -72% | 1 | 1 | 0% | 1,591 | 716 | -55% | 0 | 0 | — |
case-12 | fail→pass | 7,852 | 2,085 | -73% | 1 | 1 | 0% | 1,366 | 700 | -49% | 0 | 0 | — |
case-13 | pass→pass | 3,774 | 2,011 | -47% | 1 | 1 | 0% | 570 | 690 | +21% | 0 | 0 | — |
case-14 | pass→pass | 8,006 | 2,097 | -74% | 1 | 1 | 0% | 1,394 | 691 | -50% | 0 | 0 | — |
case-15 | pass→pass | 4,912 | 2,020 | -59% | 1 | 1 | 0% | 794 | 680 | -14% | 0 | 0 | — |
case-18 | fail→pass | 5,380 | 1,960 | -64% | 1 | 1 | 0% | 884 | 675 | -24% | 0 | 0 | — |
case-19 | pass→pass | 4,365 | 1,685 | -61% | 1 | 1 | 0% | 698 | 623 | -11% | 0 | 0 | — |
case-20 | fail→pass | 6,368 | 1,516 | -76% | 1 | 1 | 0% | 1,033 | 629 | -39% | 0 | 0 | — |
case-21 | pass→pass | 3,519 | 1,562 | -56% | 1 | 1 | 0% | 619 | 646 | +4% | 0 | 0 | — |
case-22 | pass→pass | 8,043 | 2,327 | -71% | 1 | 1 | 0% | 1,255 | 800 | -36% | 0 | 0 | — |
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.
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.
Other measured skills in the registry, with their headline benchmark lift.