Skip to main content
ClaudeWave
Skill490 repo starsupdated 3d ago

trl-training

Train and fine-tune transformer language models using TRL (Transformers Reinforcement Learning). Supports SFT, DPO, GRPO, KTO, RLOO and Reward Model training via CLI commands.

Install in Claude Code
Copy
git clone --depth 1 https://github.com/waybarrios/opencode-power-pack /tmp/trl-training && cp -r /tmp/trl-training/skills/trl-training ~/.claude/skills/trl-training
Then start a new Claude Code session; the skill loads automatically.

SKILL.md

# TRL Training Skill

You are an expert at using the TRL (Transformers Reinforcement Learning) library to train and fine-tune large language models.

## Overview

TRL provides CLI commands for post-training foundation models using state-of-the-art techniques:

- **SFT** (Supervised Fine-Tuning): Fine-tune models on instruction-following or conversational datasets
- **DPO** (Direct Preference Optimization): Align models using preference data
- **GRPO** (Group Relative Policy Optimization): Train models by ranking multiple sampled outputs relative to each other and optimizing based on their comparative rewards.
- **RLOO** (Reinforce Leave One Out): Online RL training with generation-based rewards
- **Reward Model Training**: Train reward models for RLHF

TRL is built on top of Hugging Face Transformers and Accelerate, providing seamless integration with the Hugging Face ecosystem.

## Core Commands

### trl sft - Supervised Fine-Tuning

Fine-tune language models on instruction-following or conversational datasets.

**Full training:**

```bash
trl sft \
  --model_name_or_path Qwen/Qwen2-0.5B \
  --dataset_name trl-lib/Capybara \
  --learning_rate 2.0e-5 \
  --num_train_epochs 1 \
  --packing \
  --per_device_train_batch_size 2 \
  --gradient_accumulation_steps 8 \
  --eos_token '<|im_end|>' \
  --eval_strategy steps \
  --eval_steps 100 \
  --output_dir Qwen2-0.5B-SFT \
  --push_to_hub
```

**Train with LoRA adapters:**

```bash
trl sft \
  --model_name_or_path Qwen/Qwen2-0.5B \
  --dataset_name trl-lib/Capybara \
  --learning_rate 2.0e-4 \
  --num_train_epochs 1 \
  --packing \
  --per_device_train_batch_size 2 \
  --gradient_accumulation_steps 8 \
  --eos_token '<|im_end|>' \
  --eval_strategy steps \
  --eval_steps 100 \
  --use_peft \
  --lora_r 32 \
  --lora_alpha 16 \
  --output_dir Qwen2-0.5B-SFT \
  --push_to_hub
```

### trl dpo - Direct Preference Optimization

Align models using preference data (chosen/rejected pairs).

**Full training:**

```bash
trl dpo \
  --dataset_name trl-lib/ultrafeedback_binarized \
  --model_name_or_path Qwen/Qwen2-0.5B-Instruct \
  --learning_rate 5.0e-7 \
  --num_train_epochs 1 \
  --per_device_train_batch_size 2 \
  --max_steps 1000 \
  --gradient_accumulation_steps 8 \
  --eval_strategy steps \
  --eval_steps 50 \
  --output_dir Qwen2-0.5B-DPO \
  --no_remove_unused_columns
```

**Train with LoRA adapters:**

```bash
trl dpo \
  --dataset_name trl-lib/ultrafeedback_binarized \
  --model_name_or_path Qwen/Qwen2-0.5B-Instruct \
  --learning_rate 5.0e-6 \
  --num_train_epochs 1 \
  --per_device_train_batch_size 2 \
  --max_steps 1000 \
  --gradient_accumulation_steps 8 \
  --eval_strategy steps \
  --eval_steps 50 \
  --output_dir Qwen2-0.5B-DPO \
  --no_remove_unused_columns \
  --use_peft \
  --lora_r 32 \
  --lora_alpha 16
```

### trl grpo - Group Relative Policy Optimization

Train models using reward functions or LLM-as-a-judge for evaluating generations and providing rewards.

**Basic usage:**

```bash
trl grpo \
  --model_name_or_path Qwen/Qwen2.5-0.5B \
  --dataset_name trl-lib/gsm8k \
  --reward_funcs accuracy_reward \
  --output_dir Qwen2-0.5B-GRPO \
  --push_to_hub
```

### trl rloo - Reinforce Leave One Out

Online RL training where the model generates text and receives rewards based on custom criteria.

**Basic usage:**

```bash
trl rloo \
  --model_name_or_path Qwen/Qwen2.5-0.5B \
  --dataset_name trl-lib/tldr \
  --reward_model_name_or_path sentiment-analysis:nlptown/bert-base-multilingual-uncased-sentiment \
  --output_dir Qwen2-0.5B-RLOO \
  --push_to_hub
```

### trl reward - Reward Model Training

Train a reward model to score text quality for RLHF.

**Full training:**

```bash
trl reward \
  --model_name_or_path Qwen/Qwen2-0.5B-Instruct \
  --dataset_name trl-lib/ultrafeedback_binarized \
  --output_dir Qwen2-0.5B-Reward \
  --per_device_train_batch_size 8 \
  --num_train_epochs 1 \
  --learning_rate 1.0e-5 \
  --eval_strategy steps \
  --eval_steps 50 \
  --max_length 2048
```

**Train with LoRA adapters:**

```bash
trl reward \
  --model_name_or_path Qwen/Qwen2-0.5B-Instruct \
  --dataset_name trl-lib/ultrafeedback_binarized \
  --output_dir Qwen2-0.5B-Reward-LoRA \
  --per_device_train_batch_size 8 \
  --num_train_epochs 1 \
  --learning_rate 1.0e-4 \
  --eval_strategy steps \
  --eval_steps 50 \
  --max_length 2048 \
  --use_peft \
  --lora_task_type SEQ_CLS \
  --lora_r 32 \
  --lora_alpha 16
```

## Configuration Files

TRL supports YAML configuration files for reproducible training. All CLI arguments can be specified in a config file.

**Example config (sft_config.yaml):**

```yaml
model_name_or_path: Qwen/Qwen2.5-0.5B
dataset_name: trl-lib/Capybara
learning_rate: 2.0e-5
num_train_epochs: 1
per_device_train_batch_size: 8
gradient_accumulation_steps: 2
output_dir: ./sft_output
use_peft: true
lora_r: 16
lora_alpha: 16
report_to: trackio
```

**Launch with config:**

```bash
trl sft --config sft_config.yaml
```

**Override config values:**

```bash
trl sft --config sft_config.yaml --learning_rate 1.0e-5
```

## Distributed Training

TRL integrates with Accelerate for multi-GPU and multi-node training.

**Multi-GPU training:**

```bash
trl sft \
  --config sft_config.yaml \
  --num_processes 4
```

**Use predefined Accelerate configs:**

TRL provides predefined configs: `single_gpu`, `multi_gpu`, `fsdp1`, `fsdp2`, `zero1`, `zero2`, `zero3`

```bash
trl sft \
  --config sft_config.yaml \
  --accelerate_config zero2
```

**Custom Accelerate config:**

```bash
# Generate custom config
accelerate config

# Use custom config
trl sft --config sft_config.yaml --config_file ~/.cache/huggingface/accelerate/default_config.yaml
```

**Fully Sharded Data Parallel (FSDP):**

```bash
trl sft --config sft_config.yaml --accelerate_config fsdp2
```

**DeepSpeed ZeRO:**

```bash
trl sft --config sft_config.yaml --accelerate_config zero3
```

## Troubleshooting

### CUDA Out of Memory

- Reduce `--per_device_train_batch_size` and increase `--gradi
agents-md-improverSkill

Audit and improve project-rules files (AGENTS.md, CLAUDE.md, .agents/instructions, local overrides) so the agent keeps accurate project context. Use when the user asks to check, audit, review, update, improve, or fix their AGENTS.md or CLAUDE.md, mentions "project rules maintenance" or "agent context optimization", or when the codebase has changed enough that the rules file may be stale. Scans the repository for every rules file, grades each against a quality rubric, outputs a quality report, and applies targeted edits only after user approval.

agents-md-reviseSkill

Capture learnings from the current session into the project-rules file (AGENTS.md, CLAUDE.md, or local override) so future sessions benefit. Use when the user says "revise the rules", "update AGENTS.md / CLAUDE.md with what we just learned", "save this to project memory", "remember this for next time", or at the end of a productive session when valuable context has emerged that is not yet documented. This complements agents-md-improver — improver audits, while this one captures.

code-architectSkill

Design a feature architecture by analyzing existing codebase patterns and conventions, then provide a comprehensive implementation blueprint with specific files to create or modify, component designs, data flows, and a build sequence. Use this skill when the user asks for an architecture design, an implementation plan for a non-trivial feature, or when dispatched as a sub-task during feature-dev architecture phase.

code-explorerSkill

Deeply analyze an existing codebase feature by tracing execution paths, mapping architecture layers, understanding patterns and abstractions, and documenting dependencies. Use this skill when you need to understand how a feature works before modifying or extending it, when dispatched as a sub-task during feature-dev exploration, or when the user asks "how does X work in this codebase".

code-reviewSkill

Review a pull request or a set of code changes for bugs, logic errors, and project-convention violations using a confidence-filtered, multi-agent process. Use this skill when the user asks to review a PR, audit pending changes, or inspect a diff for problems before merging.

code-reviewerSkill

Review code for bugs, logic errors, security vulnerabilities, code quality issues, and adherence to project conventions, using confidence-based filtering to report only high-priority issues that truly matter. Use this skill when reviewing a small set of changes locally (such as unstaged diff), when dispatched as a sub-task during feature-dev quality review, or when the user wants a critique of a specific file or function.

feature-devSkill

Guide a feature implementation through a structured seven-phase workflow with deep codebase understanding, clarifying questions, parallel architecture design, and quality review. Use this skill when the user asks to build a new feature, add functionality, or wants a methodical approach to implementation rather than diving straight to code.

frontend-designSkill

Create distinctive, production-grade frontend interfaces with high design quality and accessible markup. Use this skill when the user asks to build or beautify web components, pages, applications, landing pages, dashboards, artifacts, or React/HTML/CSS UI. Generates creative, polished code that avoids generic AI aesthetics, then self-checks it against an objective accessibility and quality rubric.