Unsloth Finetuning

Unsloth Fine-Tuning

Overview

Unsloth trains LLMs with custom kernels that cut VRAM use and step time without changing the math, which makes single-GPU fine-tuning practical on hardware that would otherwise OOM. This skill covers Unsloth Core — the Python API — because that is what an agent can drive programmatically; the Desktop app and Studio web UI are interactive and out of scope.

The hard parts of an Unsloth run are not the training call. They are sizing the job against available VRAM, getting the chat template and loss masking right, and choosing an export format the target runtime can actually load. This skill covers those three.

When to Use This Skill

• Use when fine-tuning an LLM on one GPU and VRAM is the binding constraint.
• Use when a training run OOMs and needs to be resized rather than rewritten.
• Use when doing preference or RL post-training (GRPO, DPO) on consumer hardware.
• Use when a fine-tuned model must be exported to GGUF, vLLM, or merged 16-bit weights.
• Use when a fine-tune "ran fine" but the model's output format is wrong — usually a chat template or loss-masking bug, not a hyperparameter one.

Do not use this skill when

• The training is multi-node or large-scale multi-GPU. Use plain TRL with Accelerate/DeepSpeed.
• The architecture is unsupported by Unsloth. Fall back to TRL; do not force it.
• The user wants managed cloud training. That is Hugging Face Jobs, not local Unsloth.
• The user wants the Desktop or Studio GUI. Point them at the installer, not this skill.

How It Works

Step 1: Size the run before writing code

VRAM is the constraint that decides everything else. Estimate weights first, then leave room for activations and optimizer state:

Add roughly 2-6 GB for activations, scaling with max_seq_length and batch size. Treat these as planning figures and confirm against nvidia-smi on the first run — they vary by architecture, attention implementation and vocabulary size.

If the estimate does not fit, reduce in this order: max_seq_length, then batch size (raising gradient_accumulation_steps to hold the effective batch constant), then LoRA rank, then model size. Cutting rank before sequence length usually costs more quality than it saves memory.

Step 2: Load the model

import unsloth must come before transformers, trl or peft. Unsloth patches those libraries at import time; importing them first silently disables the optimizations.

import unsloth  # must be first
import os
import re
from unsloth import FastLanguageModel

def reviewed_revision(variable):
    revision = os.environ.get(variable, "")
    if re.fullmatch(r"[0-9a-fA-F]{40}", revision) is None:
        raise RuntimeError(f"{variable} must be a reviewed full 40-character Hub commit SHA")
    return revision.lower()

model_revision = reviewed_revision("UNSLOTH_MODEL_REVISION")
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "unsloth/Qwen3-8B",
    revision = model_revision,
    max_seq_length = 2048,
    load_in_4bit = True,
    dtype = None,  # auto-detects bf16 where supported
)

Pick the loader that matches the modality: FastLanguageModel for text-only causal LMs, FastVisionModel for vision-language models, FastModel when the modality is decided at runtime.

The unsloth/ Hub namespace holds pre-quantized copies that download faster and skip a local quantization pass. Upstream repos such as Qwen/ or meta-llama/ work identically. Before setting UNSLOTH_MODEL_REVISION, inspect that exact Hub commit and obtain approval for the download. Record the repository and full revision with the run; never substitute a branch, tag, range, or moving default.

Step 3: Fix the chat template before training

This is the most common silent failure. A run with the wrong template converges cleanly and produces a model that ignores its stop tokens or emits prompt scaffolding at inference.

from unsloth.chat_templates import (
    get_chat_template,
    standardize_data_formats,
    train_on_responses_only,
)

tokenizer = get_chat_template(tokenizer, chat_template = "qwen3")
dataset = standardize_data_formats(dataset)  # normalizes ShareGPT/OpenAI column names

Then mask the prompt so loss is computed on assistant turns only. Without this, the model is also trained to generate user messages:

trainer = train_on_responses_only(
    trainer,
    instruction_part = "<|im_start|>user\n",
    response_part = "<|im_start|>assistant\n",
)

The two part strings must match the template's actual delimiters. Verify by decoding one batch and confirming the masked region covers exactly the prompt.

Step 4: Attach LoRA adapters

model = FastLanguageModel.get_peft_model(
    model,
    r = 16,
    lora_alpha = 16,
    lora_dropout = 0.0,
    target_modules = [
        "q_proj", "k_proj", "v_proj", "o_proj",
        "gate_proj", "up_proj", "down_proj",
    ],
    use_gradient_checkpointing = "unsloth",  # Unsloth's variant, lower VRAM than True
    random_state = 3407,
)

Rank guidance: r=8-16 for style and format adaptation, r=32-64 when teaching genuinely new capability. Setting lora_alpha to 1-2x r is a safe default. Keep lora_dropout = 0.0 — Unsloth's fast path is only taken when dropout is zero.

Train all seven projection modules unless VRAM forces otherwise; attention-only LoRA underperforms noticeably on instruction data. For MoE models, expert layers are nn.Parameter rather than nn.Linear and need target_parameters instead of target_modules.

Step 5: Train

Unsloth returns standard PEFT-wrapped models, so TRL's trainers work unmodified.

from trl import SFTTrainer, SFTConfig

trainer = SFTTrainer(
    model = model,
    tokenizer = tokenizer,
    train_dataset = dataset,
    args = SFTConfig(
        per_device_train_batch_size = 2,
        gradient_accumulation_steps = 8,  # effective batch 16
        warmup_steps = 5,
        num_train_epochs = 1,
        learning_rate = 2e-4,
        optim = "adamw_8bit",
        output_dir = "outputs",
    ),
)
trainer.train()

2e-4 suits LoRA; full fine-tuning needs roughly 10x lower. One to three epochs is typical — LoRA overfits small datasets quickly, so watch eval loss rather than trusting an epoch count.

Step 6: Export to the target runtime

The right format depends entirely on where the model will run:

quantization_method accepts a list, so several GGUF quants can be produced in one conversion pass: ["q4_k_m", "q5_k_m", "q8_0"]. q4_k_m is the usual quality/size compromise. The iq* importance-matrix quants additionally require imatrix_file=.

Avoid save_method="merged_4bit" for anything redistributed — it bakes in the quantization and cannot be cleanly re-quantized afterwards.

Examples

Example 1: QLoRA SFT on a 16 GB GPU

import unsloth
import os
import re
from unsloth import FastLanguageModel
from unsloth.chat_templates import get_chat_template, train_on_responses_only
from datasets import load_dataset
from trl import SFTTrainer, SFTConfig

def reviewed_revision(variable):
    revision = os.environ.get(variable, "")
    if re.fullmatch(r"[0-9a-fA-F]{40}", revision) is None:
        raise RuntimeError(f"{variable} must be a reviewed full 40-character Hub commit SHA")
    return revision.lower()

model_revision = reviewed_revision("UNSLOTH_MODEL_REVISION")
dataset_revision = reviewed_revision("UNSLOTH_DATASET_REVISION")
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "unsloth/Qwen3-8B",
    revision = model_revision,
    max_seq_length = 2048,
    load_in_4bit = True,
)
model = FastLanguageModel.get_peft_model(model, r = 16, lora_alpha = 16)

tokenizer = get_chat_template(tokenizer, chat_template = "qwen3")
dataset = load_dataset(
    "mlabonne/FineTome-100k",
    revision = dataset_revision,
    split = "train[:5000]",
)

trainer = SFTTrainer(
    model = model,
    tokenizer = tokenizer,
    train_dataset = dataset,
    args = SFTConfig(
        per_device_train_batch_size = 2,
        gradient_accumulation_steps = 8,
        num_train_epochs = 1,
        learning_rate = 2e-4,
        optim = "adamw_8bit",
        output_dir = "outputs",
    ),
)
trainer = train_on_responses_only(
    trainer,
    instruction_part = "<|im_start|>user\n",
    response_part = "<|im_start|>assistant\n",
)
trainer.train()

model.save_pretrained_gguf("qwen3-tuned", tokenizer, quantization_method = "q4_k_m")

Example 2: GRPO with vLLM-backed generation

GRPO samples several completions per prompt at every step, so generation dominates step time. Load with fast_inference=True to route sampling through vLLM in the same process.

import unsloth
import os
import re
from unsloth import FastLanguageModel
from trl import GRPOTrainer, GRPOConfig

def reviewed_revision(variable):
    revision = os.environ.get(variable, "")
    if re.fullmatch(r"[0-9a-fA-F]{40}", revision) is None:
        raise RuntimeError(f"{variable} must be a reviewed full 40-character Hub commit SHA")
    return revision.lower()

model_revision = reviewed_revision("UNSLOTH_MODEL_REVISION")
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "unsloth/Qwen3-4B",
    revision = model_revision,
    max_seq_length = 1024,
    load_in_4bit = True,
    fast_inference = True,       # vLLM sampling backend
    max_lora_rank = 32,          # must be >= the r used below
    gpu_memory_utilization = 0.6,
)
model = FastLanguageModel.get_peft_model(model, r = 32, lora_alpha = 32)

def reward_length(completions, **kwargs):
    """Placeholder. Replace with a task-specific verifier."""
    return [min(len(c) / 200.0, 1.0) for c in completions]

trainer = GRPOTrainer(
    model = model,
    processing_class = tokenizer,
    reward_funcs = [reward_length],
    train_dataset = dataset,
    args = GRPOConfig(
        num_generations = 8,
        max_prompt_length = 256,
        max_completion_length = 512,
        learning_rate = 5e-6,
        output_dir = "grpo-outputs",
    ),
)
trainer.train()

gpu_memory_utilization splits VRAM between vLLM's KV cache and training. Raise it if generation is the bottleneck, lower it if training OOMs. max_lora_rank is fixed at load time and must be at least the r passed later, or adapter loading fails.

GRPO learning rates sit roughly two orders of magnitude below SFT. Reward functions receive completions plus any dataset columns as keyword arguments, and return one float per completion.