Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx skills add hamzabellouch/agent-skills --skill local-model-finetuning-unsloth-axolotlgit clone --depth 1 https://github.com/hamzabellouch/agent-skillsWrote this? Show the measurements
A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.
[](https://agentmods.dev/skills/hamzabellouch/agent-skills/local-model-finetuning-unsloth-axolotl)<a href="https://agentmods.dev/skills/hamzabellouch/agent-skills/local-model-finetuning-unsloth-axolotl"><img src="https://agentmods.dev/badge/skills/hamzabellouch/agent-skills/local-model-finetuning-unsloth-axolotl/github.svg" alt="Measured on agentmods" height="20"></a>Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.
<a href="https://agentmods.dev/skills/hamzabellouch/agent-skills/local-model-finetuning-unsloth-axolotl"><img src="https://agentmods.dev/badge/skills/hamzabellouch/agent-skills/local-model-finetuning-unsloth-axolotl.svg" alt="Reviewed on agentmods" width="80" height="20"></a>What it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5.1 | $0.00085 | $0.02141 |
| Opus 5 | $0.00043 | $0.01071 |
| Sonnet 5 | $0.00017 | $0.00428 |
| Haiku 4.5 | $0.00009 | $0.00214 |
Grade A, and why
local-model-finetuning-unsloth-axolotl scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 12d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 193 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Local LLM Fine-Tuning Architect Skill: Unsloth & Axolotl
1. Framework Architectural Comparison
| Dimension | Unsloth | Axolotl |
|---|---|---|
| Primary Target | Single-GPU extreme speed & VRAM optimization | Multi-GPU / Multi-Node enterprise scale |
| Backend Implementation | Custom C++/CUDA & Triton kernels (manual backprop) | HuggingFace Transformers, PyTorch FSDP, DeepSpeed |
| Interface | Python API (Extends trl & peft) |
YAML Configuration Driven CLI |
| Speedup vs Standard | 2x – 5x faster training | Standard PyTorch + FlashAttention-2 optimizations |
| Memory Footprint | Up to 80% VRAM reduction | Standard QLoRA/LoRA VRAM scaling |
| Alignment Algorithms | SFT, DPO, ORPO, GRPO | SFT, DPO, ORPO, KTO, PPO, ReFT |
| Model Architectures | Llama 3/3.1/3.2, Qwen 2.5, Mistral, Gemma 2, Phi-4 | Broad HF ecosystem support (Llama, Qwen, Mistral, etc.) |
2. Memory Optimization & Hardware Configurations
VRAM Budgeting Matrix (8B Model @ 4096 Sequence Length)
| Method | Quantization | Batch Size (per GPU) | Min VRAM Required | Optimal Hardware |
|---|---|---|---|---|
| Unsloth QLoRA | 4-bit (NF4) | 2 – 4 | 7 GB – 10 GB | RTX 3090 / RTX 4090 / A10G |
| Unsloth LoRA | 16-bit (BF16) | 1 – 2 | 16 GB – 20 GB | RTX 4090 / A100 (40GB) |
| Axolotl QLoRA (FSDP) | 4-bit (NF4) | 4 – 8 (across 4 GPUs) | 12 GB per GPU | 4x RTX 3090 / 4x A10G |
| Axolotl Full Params (DeepSpeed Z3) | 16-bit (BF16) | 2 – 4 (across 8 GPUs) | 40 GB per GPU | 8x A100 (80GB) / H100 |
Key Optimization Knobs
- NF4 & Double Quantization: Uses 4-bit NormalFloat data type with quantized quantization constants to save ~0.5 bit per parameter.
- Paged AdamW 8-bit: Offloads optimizer state spikes to CPU memory during peak backpropagation passes.
- Gradient Checkpointing (Unsloth Offloading): Recomputes activations during backpass instead of storing them all in RAM. Unsloth reduces activation memory footprint by 50-70%.
- Sample Packing / Multipack: Concatenates short samples into a single sequence up to max token length, eliminating padding token waste and accelerating training by 2x-4x.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 12d ago First seen · 193 lines · 85 tokens per session scan A 37c24822aaa4
local-model-finetuning-unsloth-axolotl is a skill published in the GitHub repository hamzabellouch/agent-skills (4 stars, last pushed 1mo ago), licensed MIT. It adds 85 tokens to every session and 2,141 once invoked, about $0.0004 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
prompt-engineering
Universal prompt engineering techniques for any LLM. Use when crafting, optimizing, or reviewing prompts for AI models. Triggers on requests like "improve this prompt", "write a system prompt", "optimize my instructions", "help me prompt engineer", "audit this prompt", "review my prompt", or when building agentic…
fixing-prompt
Prompt: Prompt Refinement and Optimization.
feature-engineering
When building training datasets, designing feature pipelines, or debugging training-serving skew.
streaming-patterns
When designing Kafka consumers/producers or implementing real-time pipelines.
inference-serving
When deploying a model to an API endpoint or optimizing inference latency.
model-evaluation
When evaluating a trained model, comparing versions, or performing fairness analysis.