Unsloth Training

1. GRPO - RL with reward functions (no labeled outputs needed)
2. SFT - Supervised fine-tuning with input/output pairs
3. Vision - VLM fine-tuning (Qwen3-VL, Gemma3, Llama 3.2 Vision)

Key capabilities:

• FP8 Training - 60% less VRAM, 1.4x faster (RTX 40+, H100)
• 3x Packing - Automatic 2-5x speedup for mixed-length data
• Docker - Official unsloth/unsloth image
• Mobile - QAT → ExecuTorch → iOS/Android (~40 tok/s)
• Export - GGUF, Ollama, vLLM, LM Studio, SGLang

model, tokenizer = FastLanguageModel.from_pretrained( model_name="unsloth/Qwen3-8B", max_seq_length=2048, load_in_fp8=True, fast_inference=True, ) model = FastLanguageModel.get_peft_model( model, r=64, target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"], use_gradient_checkpointing="unsloth", )

def correctness_reward(completions, answer, **kwargs): return [2.0 if extract_answer(c) == a else 0.0 for c, a in zip(completions, answer)]

trainer = GRPOTrainer( model=model, args=GRPOConfig(num_generations=4, beta=0.04, learning_rate=5e-6), train_dataset=dataset, reward_funcs=[correctness_reward], ) trainer.train()

**SFT with Packing (2-5x faster):**
```python
from trl import SFTTrainer, SFTConfig

trainer = SFTTrainer(
    model=model, train_dataset=dataset, processing_class=tokenizer,
    args=SFTConfig(
        per_device_train_batch_size=2, num_train_epochs=3,
        learning_rate=2e-4, packing=True,  # 2-5x speedup
    ),
)
trainer.train()

Natural language:

• "train with GRPO", "fine-tune", "reward functions"
• "FP8 training", "fp8", "less VRAM"
• "vision fine-tuning", "VLM", "image training"
• "phone deployment", "mobile LLM", "ExecuTorch"
• "docker training", "container", "unsloth docker"
• "packing", "faster training", "500k context"
• "export GGUF", "Ollama", "vLLM", "SGLang"
• "install unsloth", "pip install", "setup unsloth"
• "prepare dataset", "training data", "synthetic data", "ChatML", "ShareGPT"
• "environment flags", "UNSLOTH_RETURN_LOGITS"
• "MLX", "Apple Silicon", "Mac training", "unsloth-mlx"

Setup and data references:

• reference/installation.md - pip/uv install, CUDA versions, venv, Colab
• reference/environment-flags.md - UNSLOTH_RETURN_LOGITS, COMPILE_DISABLE, etc.
• reference/datasets-guide.md - Formats (ChatML/ShareGPT/Alpaca), chat templates, synthetic data
• reference/mlx-training.md - Apple Silicon training with unsloth-mlx

Training feature references:

• reference/fp8-training.md - FP8 setup, VRAM savings
• reference/deployment.md - Docker, vLLM, LoRA hot-swap, SGLang
• reference/export-formats.md - GGUF, Ollama, LM Studio, Dynamic 2.0
• reference/advanced-training.md - 500K context, packing, checkpoints
• reference/vision-training.md - VLM fine-tuning
• reference/mobile-deployment.md - QAT, ExecuTorch, iOS/Android

Code examples: reference/grpo/, reference/sft/

Model Selection

Core Hyperparameters

GRPO (RL):

GRPOConfig(
    num_generations=4,        # Completions per prompt (2-8)
    beta=0.04,                # KL penalty (0.01-0.1)
    learning_rate=5e-6,       # 10x smaller than SFT!
    max_completion_length=512,
    max_steps=300,            # Minimum for results
)

SFT:

TrainingArguments(
    learning_rate=2e-4,       # Standard SFT rate
    num_train_epochs=3,       # 2-4 typical
    per_device_train_batch_size=2,
)

Reward functions are the core of GRPO. They return a list of floats for each completion.

Pattern 1: Correctness (Primary Signal)

def correctness_reward(completions, answer, **kwargs):
    """
    +2.0 for correct answer, 0.0 otherwise.
    This should be your highest-weighted reward.
    """
    rewards = []
    for completion, true_answer in zip(completions, answer):
        extracted = extract_answer(completion)
        try:
            pred = float(extracted.replace(",", "").strip())
            true = float(true_answer.replace(",", "").strip())
            reward = 2.0 if abs(pred - true) < 0.01 else 0.0
        except ValueError:
            reward = 2.0 if extracted.strip() == str(true_answer).strip() else 0.0
        rewards.append(reward)
    return rewards

Pattern 2: Format Compliance

def format_reward(completions, **kwargs):
    """
    +0.5 for proper XML structure with reasoning and answer tags.
    """
    rewards = []
    for completion in completions:
        has_reasoning = bool(re.search(r"<reasoning>.*?</reasoning>", completion, re.DOTALL))
        has_answer = bool(re.search(r"<answer>.*?</answer>", completion, re.DOTALL))
        if has_reasoning and has_answer:
            rewards.append(0.5)
        elif has_answer:
            rewards.append(0.2)
        else:
            rewards.append(0.0)
    return rewards

Pattern 3: Reasoning Quality

def reasoning_length_reward(completions, **kwargs):
    """
    +0.3 for substantive reasoning (30-200 words).
    """
    rewards = []
    for completion in completions:
        reasoning = extract_reasoning(completion)
        word_count = len(reasoning.split()) if reasoning else 0
        if 30 <= word_count <= 200:
            rewards.append(0.3)
        elif 15 <= word_count < 30:
            rewards.append(0.1)
        else:
            rewards.append(0.0)
    return rewards

Pattern 4: Negative Constraints

def no_hedging_reward(completions, **kwargs):
    """
    -0.3 penalty for uncertainty language.
    """
    hedging = ["i think", "maybe", "perhaps", "possibly", "i'm not sure"]
    rewards = []
    for completion in completions:
        has_hedging = any(phrase in completion.lower() for phrase in hedging)
        rewards.append(-0.3 if has_hedging else 0.0)
    return rewards

Typical Reward Stack

reward_funcs = [
    correctness_reward,      # +2.0 max (primary signal)
    format_reward,           # +0.5 max (structure)
    reasoning_length_reward, # +0.3 max (quality)
    no_hedging_reward,       # -0.3 max (constraint)
]
# Total range: -0.3 to +2.8

For domain-specific rewards: See reference/domain-examples.md for Voice AI, Sales Agent, and Support patterns.

System Prompt with XML Tags

SYSTEM_PROMPT = """You are a helpful assistant that thinks step-by-step.

Always respond in this exact format:
<reasoning>
[Your step-by-step thinking process]
</reasoning>
<answer>
[Your final answer - just the number or short response]
</answer>
"""

Extraction Helpers

import re

def extract_answer(text: str) -> str:
    """Extract answer from XML tags"""
    match = re.search(r"<answer>(.*?)</answer>", text, re.DOTALL)
    return match.group(1).strip() if match else ""

def extract_reasoning(text: str) -> str:
    """Extract reasoning from XML tags"""
    match = re.search(r"<reasoning>(.*?)</reasoning>", text, re.DOTALL)
    return match.group(1).strip() if match else ""

Dataset Format

GRPO (prompt-only):

dataset = dataset.map(lambda ex: {
    "prompt": [
        {"role": "system", "content": SYSTEM_PROMPT},
        {"role": "user", "content": ex["question"]}
    ],
    "answer": ex["answer"]  # Ground truth for verification
})

SFT (full conversations):

dataset = dataset.map(lambda ex: {
    "conversations": [
        {"role": "system", "content": SYSTEM_PROMPT},
        {"role": "user", "content": ex["input"]},
        {"role": "assistant", "content": ex["output"]}
    ]
})

Save LoRA Only (~100MB)

model.save_lora("grpo_lora")

Merge and Save Full Model

model.save_pretrained_merged(
    "grpo_merged", tokenizer,
    save_method="merged_16bit",
)

Export to GGUF for Ollama

model.save_pretrained_gguf(
    "grpo_gguf", tokenizer,
    quantization_method="q4_k_m",  # Options: q4_k_m, q8_0, q5_k_m
)

Test with Ollama

# Create Modelfile
cat > Modelfile << EOF
FROM ./grpo_gguf/unsloth.Q4_K_M.gguf
TEMPLATE """{{ .System }}
User: {{ .Prompt }}
Assistant: """
PARAMETER temperature 0.7
EOF

ollama create my-model -f Modelfile
ollama run my-model "Solve: 15 + 27 = ?"

GRPO training: → GRPOConfig, reward functions, dataset prep → Reference: reference/grpo/basic_grpo.py

SFT training: → SFTTrainer, dataset formatting → Reference: reference/sft/sales_extractor_training.py

Reward function design: → 4 patterns (correctness, format, quality, constraints) → Reference: reference/reward-design.md, reference/domain-examples.md

FP8 training: → 60% VRAM savings, env vars, pre-quantized models → Reference: reference/fp8-training.md

Docker setup: → Official image, volumes, Jupyter/SSH → Reference: reference/deployment.md

Vision fine-tuning: → FastVisionModel, VLM data format → Reference: reference/vision-training.md

Mobile deployment: → QAT, ExecuTorch, iOS/Android → Reference: reference/mobile-deployment.md

Long context / packing: → 500K context, 2-5x speedup → Reference: reference/advanced-training.md

Export formats: → GGUF methods, Ollama, vLLM, SGLang → Reference: reference/export-formats.md

Installation: → pip, uv, Colab, CUDA versions → Reference: reference/installation.md

Dataset preparation: → Formats, chat templates, synthetic data, multi-column → Reference: reference/datasets-guide.md

Environment flags: → UNSLOTH_RETURN_LOGITS, COMPILE_DISABLE, etc. → Reference: reference/environment-flags.md

Apple Silicon / MLX: → unsloth-mlx, Mac training → Reference: reference/mlx-training.md

Training issues: → reference/troubleshooting.md

For detailed troubleshooting: See reference/troubleshooting.md

GRPO:

• Model loads without OOM
• LoRA configured with use_gradient_checkpointing="unsloth"
• Dataset has prompt and answer fields
• At least one reward function defined and tested
• num_generations >= 2
• beta set (0.01-0.1, start at 0.04)
• learning_rate set (1e-6 to 1e-5)
• At least 300 steps planned

SFT:

• Model loads without OOM
• Dataset has conversations or text field
• Chat template applied correctly
• learning_rate ~2e-4
• 2-4 epochs planned

Training Data Generation:

• 50 seeds x 10 variations x ~1000 tokens = ~500K tokens
• Claude Sonnet: ~$1.50
• GPT-4o-mini: ~$0.30

Training Compute:

• Google Colab T4: Free (30-60 min for 1000 examples)
• RunPod A40 (48GB): ~$0.79/hr

Total per experiment: ~$0.50-$1.00 on RunPod

User: "I want to train a model that gives concise voice responses"

Claude:

1. First, let's set up the model:

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="Qwen/Qwen2.5-3B-Instruct",
    max_seq_length=1024, load_in_4bit=True,
)

2. For voice AI, you need brevity rewards. Here's a custom stack:

def brevity_reward(completions, **kwargs):
    """Voice responses under 50 words get +0.5"""
    return [0.5 if len(c.split()) <= 50 else -0.3 for c in completions]

def speakable_reward(completions, **kwargs):
    """Penalize markdown that sounds bad spoken"""
    bad = ["**", "- ", "```", "http"]
    return [-0.2 * sum(1 for b in bad if b in c) for c in completions]

reward_funcs = [correctness_reward, brevity_reward, speakable_reward]

3. See reference/domain-examples.md for complete Voice AI reward patterns.