{"id":"trl-fine-tuning","name":"fine-tuning-with-trl","summary":"強化学習を用いて強化学習を用いて、命令調整にはSFT、好みの整合のためのDPO、報酬最適化のためのPPO/GRPO、報酬モデルのトレーニングは利用可能です。","body":"# TRL - Transformer Reinforcement Learning\n\n## Quick start\n\nTRL provides post-training methods for aligning language models with human preferences.\n\n**Installation**:\n```bash\npip install trl transformers datasets peft accelerate\n```\n\n**Supervised Fine-Tuning** (instruction tuning):\n```python\nfrom trl import SFTTrainer\n\ntrainer = SFTTrainer(\n    model=\"Qwen/Qwen2.5-0.5B\",\n    train_dataset=dataset,  # Prompt-completion pairs\n)\ntrainer.train()\n```\n\n**DPO** (align with preferences):\n```python\nfrom trl import DPOTrainer, DPOConfig\n\nconfig = DPOConfig(output_dir=\"model-dpo\", beta=0.1)\ntrainer = DPOTrainer(\n    model=model,\n    args=config,\n    train_dataset=preference_dataset,  # chosen/rejected pairs\n    processing_class=tokenizer\n)\ntrainer.train()\n```\n\n## Common workflows\n\n### Workflow 1: Full RLHF pipeline (SFT → Reward Model → PPO)\n\nComplete pipeline from base model to human-aligned model.\n\nCopy this checklist:\n\n```\nRLHF Training:\n- [ ] Step 1: Supervised fine-tuning (SFT)\n- [ ] Step 2: Train reward model\n- [ ] Step 3: PPO reinforcement learning\n- [ ] Step 4: Evaluate aligned model\n```\n\n**Step 1: Supervised fine-tuning**\n\nTrain base model on instruction-following data:\n\n```python\nfrom transformers import AutoModelForCausalLM, AutoTokenizer\nfrom trl import SFTTrainer, SFTConfig\nfrom datasets import load_dataset\n\n# Load model\nmodel = AutoModelForCausalLM.from_pretrained(\"Qwen/Qwen2.5-0.5B\")\ntokenizer = AutoTokenizer.from_pretrained(\"Qwen/Qwen2.5-0.5B\")\n\n# Load instruction dataset\ndataset = load_dataset(\"trl-lib/Capybara\", split=\"train\")\n\n# Configure training\ntraining_args = SFTConfig(\n    output_dir=\"Qwen2.5-0.5B-SFT\",\n    per_device_train_batch_size=4,\n    num_train_epochs=1,\n    learning_rate=2e-5,\n    logging_steps=10,\n    save_strategy=\"epoch\"\n)\n\n# Train\ntrainer = SFTTrainer(\n    model=model,\n    args=training_args,\n    train_dataset=dataset,\n    tokenizer=tokenizer\n)\ntrainer.train()\ntrainer.save_model()\n```\n\n**Step 2: Train reward model**\n\nTrain model to predict human preferences:\n\n```python\nfrom transformers import AutoModelForSequenceClassification\nfrom trl import RewardTrainer, RewardConfig\n\n# Load SFT model as base\nmodel = AutoModelForSequenceClassification.from_pretrained(\n    \"Qwen2.5-0.5B-SFT\",\n    num_labels=1  # Single reward score\n)\ntokenizer = AutoTokenizer.from_pretrained(\"Qwen2.5-0.5B-SFT\")\n\n# Load preference data (chosen/rejected pairs)\ndataset = load_dataset(\"trl-lib/ultrafeedback_binarized\", split=\"train\")\n\n# Configure training\ntraining_args = RewardConfig(\n    output_dir=\"Qwen2.5-0.5B-Reward\",\n    per_device_train_batch_size=2,\n    num_train_epochs=1,\n    learning_rate=1e-5\n)\n\n# Train reward model\ntrainer = RewardTrainer(\n    model=model,\n    args=training_args,\n    processing_class=tokenizer,\n    train_dataset=dataset\n)\ntrainer.train()\ntrainer.save_model()\n```\n\n**Step 3: PPO reinforcement learning**\n\nOptimize policy using reward model:\n\n```bash\npython -m trl.scripts.ppo \\\n    --model_name_or_path Qwen2.5-0.5B-SFT \\\n    --reward_model_path Qwen2.5-0.5B-Reward \\\n    --dataset_name trl-internal-testing/descriptiveness-sentiment-trl-style \\\n    --output_dir Qwen2.5-0.5B-PPO \\\n    --learning_rate 3e-6 \\\n    --per_device_train_batch_size 64 \\\n    --total_episodes 10000\n```\n\n**Step 4: Evaluate**\n\n```python\nfrom transformers import pipeline\n\n# Load aligned model\ngenerator = pipeline(\"text-generation\", model=\"Qwen2.5-0.5B-PPO\")\n\n# Test\nprompt = \"Explain quantum computing to a 10-year-old\"\noutput = generator(prompt, max_length=200)[0][\"generated_text\"]\nprint(output)\n```\n\n### Workflow 2: Simple preference alignment with DPO\n\nAlign model with preferences without reward model.\n\nCopy this checklist:\n\n```\nDPO Training:\n- [ ] Step 1: Prepare preference dataset\n- [ ] Step 2: Configure DPO\n- [ ] Step 3: Train with DPOTrainer\n- [ ] Step 4: Evaluate alignment\n```\n\n**Step 1: Prepare preference dataset**\n\nDataset format:\n```json\n{\n  \"prompt\": \"What is the capital of France?\",\n  \"chosen\": \"The capital of France is Paris.\",\n  \"rejected\": \"I don't know.\"\n}\n```\n\nLoad dataset:\n```python\nfrom datasets import load_dataset\n\ndataset = load_dataset(\"trl-lib/ultrafeedback_binarized\", split=\"train\")\n# Or load your own\n# dataset = load_dataset(\"json\", data_files=\"preferences.json\")\n```\n\n**Step 2: Configure DPO**\n\n```python\nfrom trl import DPOConfig\n\nconfig = DPOConfig(\n    output_dir=\"Qwen2.5-0.5B-DPO\",\n    per_device_train_batch_size=4,\n    num_train_epochs=1,\n    learning_rate=5e-7,\n    beta=0.1,  # KL penalty strength\n    max_prompt_length=512,\n    max_length=1024,\n    logging_steps=10\n)\n```\n\n**Step 3: Train with DPOTrainer**\n\n```python\nfrom transformers import AutoModelForCausalLM, AutoTokenizer\nfrom trl import DPOTrainer\n\nmodel = AutoModelForCausalLM.from_pretrained(\"Qwen/Qwen2.5-0.5B-Instruct\")\ntokenizer = AutoTokenizer.from_pretrained(\"Qwen/Qwen2.5-0.5B-Instruct\")\n\ntrainer = DPOTrainer(\n    model=model,\n    args=config,\n    train_dataset=dataset,\n    processing_class=tokenizer\n)\n\ntrainer.train()\ntrainer.save_model()\n```\n\n**CLI alternative**:\n```bash\ntrl dpo \\\n    --model_name_or_path Qwen/Qwen2.5-0.5B-Instruct \\\n    --dataset_name argilla/Capybara-Preferences \\\n    --output_dir Qwen2.5-0.5B-DPO \\\n    --per_device_train_batch_size 4 \\\n    --learning_rate 5e-7 \\\n    --beta 0.1\n```\n\n### Workflow 3: Memory-efficient online RL with GRPO\n\nTrain with reinforcement learning using minimal memory.\n\nCopy this checklist:\n\n```\nGRPO Training:\n- [ ] Step 1: Define reward function\n- [ ] Step 2: Configure GRPO\n- [ ] Step 3: Train with GRPOTrainer\n```\n\n**Step 1: Define reward function**\n\n```python\ndef reward_function(completions, **kwargs):\n    \"\"\"\n    Compute rewards for completions.\n\n    Args:\n        completions: List of generated texts\n\n    Returns:\n        List of reward scores (floats)\n    \"\"\"\n    rewards = []\n    for completion in completions:\n        # Example: reward based on length and unique words\n        score = len(completion.split())  # Favor longer responses\n        score += len(set(completion.lower().split()))  # Reward unique words\n        rewards.append(score)\n    return rewards\n```\n\nOr use a reward model:\n```python\nfrom transformers import pipeline\n\nreward_model = pipeline(\"text-classification\", model=\"reward-model-path\")\n\ndef reward_from_model(completions, prompts, **kwargs):\n    # Combine prompt + completion\n    full_texts = [p + c for p, c in zip(prompts, completions)]\n    # Get reward scores\n    results = reward_model(full_texts)\n    return [r[\"score\"] for r in results]\n```\n\n**Step 2: Configure GRPO**\n\n```python\nfrom trl import GRPOConfig\n\nconfig = GRPOConfig(\n    output_dir=\"Qwen2-GRPO\",\n    per_device_train_batch_size=4,\n    num_train_epochs=1,\n    learning_rate=1e-5,\n    num_generations=4,  # Generate 4 completions per prompt\n    max_new_tokens=128\n)\n```\n\n**Step 3: Train with GRPOTrainer**\n\n```python\nfrom datasets import load_dataset\nfrom trl import GRPOTrainer\n\n# Load prompt-only dataset\ndataset = load_dataset(\"trl-lib/tldr\", split=\"train\")\n\ntrainer = GRPOTrainer(\n    model=\"Qwen/Qwen2-0.5B-Instruct\",\n    reward_funcs=reward_function,  # Your reward function\n    args=config,\n    train_dataset=dataset\n)\n\ntrainer.train()\n```\n\n**CLI**:\n```bash\ntrl grpo \\\n    --model_name_or_path Qwen/Qwen2-0.5B-Instruct \\\n    --dataset_name trl-lib/tldr \\\n    --output_dir Qwen2-GRPO \\\n    --num_generations 4\n```\n\n## When to use vs alternatives\n\n**Use TRL when:**\n- Need to align model with human preferences\n- Have preference data (chosen/rejected pairs)\n- Want to use reinforcement learning (PPO, GRPO)\n- Need reward model training\n- Doing RLHF (full pipeline)\n\n**Method selection**:\n- **SFT**: Have prompt-completion pairs, want basic instruction following\n- **DPO**: Have preferences, want simple alignment (no reward model needed)\n- **PPO**: Have reward model, need maximum control over RL\n- **GRPO**: Memory-constrained, want online RL\n- **Reward Model**: Building RLHF pipeline, need to score generations\n\n**Use alternatives instead:**\n- **HuggingFace Trainer**: Basic fine-tuning without RL\n- **Axolotl**: YAML-based training configuration\n- **LitGPT**: Educational, minimal fine-tuning\n- **Unsloth**: Fast LoRA training\n\n## Common issues\n\n**Issue: OOM during DPO training**\n\nReduce batch size and sequence length:\n```python\nconfig = DPOConfig(\n    per_device_train_batch_size=1,  # Reduce from 4\n    max_length=512,  # Reduce from 1024\n    gradient_accumulation_steps=8  # Maintain effective batch\n)\n```\n\nOr use gradient checkpointing:\n```python\nmodel.gradient_checkpointing_enable()\n```\n\n**Issue: Poor alignment quality**\n\nTune beta parameter:\n```python\n# Higher beta = more conservative (stays closer to reference)\nconfig = DPOConfig(beta=0.5)  # Default 0.1\n\n# Lower beta = more aggressive alignment\nconfig = DPOConfig(beta=0.01)\n```\n\n**Issue: Reward model not learning**\n\nCheck loss type and learning rate:\n```python\nconfig = RewardConfig(\n    learning_rate=1e-5,  # Try different LR\n    num_train_epochs=3  # Train longer\n)\n```\n\nEnsure preference dataset has clear winners:\n```python\n# Verify dataset\nprint(dataset[0])\n# Should have clear chosen > rejected\n```\n\n**Issue: PPO training unstable**\n\nAdjust KL coefficient:\n```python\nconfig = PPOConfig(\n    kl_coef=0.1,  # Increase from 0.05\n    cliprange=0.1  # Reduce from 0.2\n)\n```\n\n## Advanced topics\n\n**SFT training guide**: See [references/sft-training.md](references/sft-training.md) for dataset formats, chat templates, packing strategies, and multi-GPU training.\n\n**DPO variants**: See [references/dpo-variants.md](references/dpo-variants.md) for IPO, cDPO, RPO, and other DPO loss functions with recommended hyperparameters.\n\n**Reward modeling**: See [references/reward-modeling.md](references/reward-modeling.md) for outcome vs process rewards, Bradley-Terry loss, and reward model evaluation.\n\n**Online RL methods**: See [references/online-rl.md](references/online-rl.md) for PPO, GRPO, RLOO, and OnlineDPO with detailed configurations.\n\n## Hardware requirements\n\n- **GPU**: NVIDIA (CUDA required)\n- **VRAM**: Depends on model and method\n  - SFT 7B: 16GB (with LoRA)\n  - DPO 7B: 24GB (stores reference model)\n  - PPO 7B: 40GB (policy + reward model)\n  - GRPO 7B: 24GB (more memory efficient)\n- **Multi-GPU**: Supported via `accelerate`\n- **Mixed precision**: BF16 recommended (A100/H100)\n\n**Memory optimization**:\n- Use LoRA/QLoRA for all methods\n- Enable gradient checkpointing\n- Use smaller batch sizes with gradient accumulation\n\n## Resources\n\n- Docs: https://huggingface.co/docs/trl/\n- GitHub: https://github.com/huggingface/trl\n- Papers:\n  - \"Training language models to follow instructions with human feedback\" (InstructGPT, 2022)\n  - \"Direct Preference Optimization: Your Language Model is Secretly a Reward Model\" (DPO, 2023)\n  - \"Group Relative Policy Optimization\" (GRPO, 2024)\n- Examples: https://github.com/huggingface/trl/tree/main/examples/scripts","author":"@Orchestra-Research","ownerProfile":null,"authorContacts":null,"sourceUrl":"https://github.com/Orchestra-Research/AI-Research-SKILLs/tree/main/06-post-training/trl-fine-tuning","license":"MIT","category":"writing","lang":"en","tokens":2931,"stars":0,"calls30d":2,"claimed":false,"visibility":"public","origin":"crawler","version":"0.1.0","createdAt":"2026-08-22","updatedAt":"2026-08-22","files":[{"path":"references/dpo-variants.md","size":4297,"sha256":"c802f74f3f48fcde0e285f97eee90c0209740a953f257b462b6e952d982dc1e6"},{"path":"references/online-rl.md","size":1971,"sha256":"35e0ad19a8e6a4bc2c1a87e1ab5e2bfe31fb74ba07b1f4b161a1c258e0c44958"},{"path":"references/reward-modeling.md","size":2597,"sha256":"782edd743d66dec5efd1e7ec48fa23ad857102117a569ddda836b856f56068e3"},{"path":"references/sft-training.md","size":3237,"sha256":"2e42b55d3327caeab7611f94976bcca315ee1b68a5bf9a4ac437099bb2ab9418"}],"requires":{"mcp":[],"tools":[]},"safety":{"flags":[],"scannedAt":"2026-08-22","hasScripts":false,"networkEndpoints":["arxiv.org","huggingface.co"]}}