Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.
Quelldateien prüfen
Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Ein direkter Befehl überspringt den Prüf-Prompt. Prüfen Sie die Quelle, bevor Sie ihn ausführen.
Train language models using TRL (Transformer Reinforcement Learning) on fully managed Hugging Face infrastructure. No local GPU setup required—models train on cloud GPUs and results are automatically saved to the Hugging Face Hub.
TRL provides multiple training methods:
SFT (Supervised Fine-Tuning) - Standard instruction tuning
DPO (Direct Preference Optimization) - Alignment from preference data
GRPO (Group Relative Policy Optimization) - Online RL training
Reward Modeling - Train reward models for RLHF
For detailed TRL method documentation:
hf_doc_search("your query", product="trl")
hf_doc_fetch("https://huggingface.co/docs/trl/sft_trainer") # SFT
hf_doc_fetch("https://huggingface.co/docs/trl/dpo_trainer") # DPO# etc.
See also:references/training_methods.md for method overviews and selection guidance
When to Use This Skill
Use this skill when users want to:
Fine-tune language models on cloud GPUs without local infrastructure
Train with TRL methods (SFT, DPO, GRPO, etc.)
Run training jobs on Hugging Face Jobs infrastructure
Convert trained models to GGUF for local deployment (Ollama, LM Studio, llama.cpp)
Ensure trained models are permanently saved to the Hub
Use modern workflows with optimized defaults
When to Use Unsloth
Use Unsloth (references/unsloth.md) instead of standard TRL when:
Limited GPU memory - Unsloth uses ~60% less VRAM
Speed matters - Unsloth is ~2x faster
Training large models (>13B) - memory efficiency is critical
Training Vision-Language Models (VLMs) - Unsloth has FastVisionModel support
See references/unsloth.md for complete Unsloth documentation and scripts/unsloth_sft_example.py for a production-ready training script.
Key Directives
When assisting with training jobs:
ALWAYS use hf_jobs() MCP tool - Submit jobs using hf_jobs("uv", {...}), NOT bash trl-jobs commands. The script parameter accepts Python code directly. Do NOT save to local files unless the user explicitly requests it. Pass the script content as a string to hf_jobs(). If user asks to "train a model", "fine-tune", or similar requests, you MUST create the training script AND submit the job immediately using hf_jobs().
Always include Trackio - Every training script should include Trackio for real-time monitoring. Use example scripts in scripts/ as templates.
Provide job details after submission - After submitting, provide job ID, monitoring URL, estimated time, and note that the user can request status checks later.
Use example scripts as templates - Reference scripts/train_sft_example.py, scripts/train_dpo_example.py, etc. as starting points.
Local Script Execution
Repository scripts use PEP 723 inline dependencies. Run them with uv run:
uv run scripts/estimate_cost.py --help
uv run scripts/dataset_inspector.py --help
Prerequisites Checklist
Before starting any training job, verify:
✅ Account & Authentication
Hugging Face Account with Pro, Team, or Enterprise plan (Jobs require paid plan)
Authenticated login: Check with hf_whoami()
HF_TOKEN for Hub Push ⚠️ CRITICAL - Training environment is ephemeral, must push to Hub or ALL training results are lost
Token must have write permissions
MUST pass secrets={"HF_TOKEN": "$HF_TOKEN"} in job config to make token available (the $HF_TOKEN syntax
references your actual token value)
✅ Dataset Requirements
Dataset must exist on Hub or be loadable via datasets.load_dataset()
Format must match training method (SFT: "messages"/text/prompt-completion; DPO: chosen/rejected; GRPO: prompt-only)
ALWAYS validate unknown datasets before GPU training to prevent format failures (see Dataset Validation section below)
Size appropriate for hardware (Demo: 50-100 examples on t4-small; Production: 1K-10K+ on a10g-large/a100-large)
⚠️ Critical Settings
Timeout must exceed expected training time - Default 30min is TOO SHORT for most training. Minimum recommended: 1-2 hours. Job fails and loses all progress if timeout is exceeded.
Hub push must be enabled - Config: push_to_hub=True, hub_model_id="username/model-name"; Job: secrets={"HF_TOKEN": "$HF_TOKEN"}
Asynchronous Job Guidelines
⚠️ IMPORTANT: Training jobs run asynchronously and can take hours
Action Required
When user requests training:
Create the training script with Trackio included (use scripts/train_sft_example.py as template)
Submit immediately using hf_jobs() MCP tool with script content inline - don't save to file unless user requests
Report submission with job ID, monitoring URL, and estimated time
Wait for user to request status checks - don't poll automatically
Ground Rules
Jobs run in background - Submission returns immediately; training continues independently
Initial logs delayed - Can take 30-60 seconds for logs to appear
User checks status - Wait for user to request status updates
Avoid polling - Check logs only on user request; provide monitoring links instead
After Submission
Provide to user:
✅ Job ID and monitoring URL
✅ Expected completion time
✅ Trackio dashboard URL
✅ Note that user can request status checks later
Example Response:
✅ Job submitted successfully!
Job ID: abc123xyz
Monitor: https://huggingface.co/jobs/username/abc123xyz
Expected time: ~2 hours
Estimated cost: ~$10
The job is running in the background. Ask me to check status/logs when ready!
Quick Start: Three Approaches
💡 Tip for Demos: For quick demos on smaller GPUs (t4-small), omit eval_dataset and eval_strategy to save ~40% memory. You'll still see training loss and learning progress.
Sequence Length Configuration
TRL config classes use max_length (not max_seq_length) to control tokenized sequence length:
# ✅ CORRECT - If you need to set sequence length
SFTConfig(max_length=512) # Truncate sequences to 512 tokens
DPOConfig(max_length=2048) # Longer context (2048 tokens)# ❌ WRONG - This parameter doesn't exist
SFTConfig(max_seq_length=512) # TypeError!
Default behavior:max_length=1024 (truncates from right). This works well for most training.
When to override:
Longer context: Set higher (e.g., max_length=2048)
Memory constraints: Set lower (e.g., max_length=512)
Vision models: Set max_length=None (prevents cutting image tokens)
Usually you don't need to set this parameter at all - the examples below use the sensible default.
UV scripts use PEP 723 inline dependencies for clean, self-contained training. This is the primary approach for Claude Code.
hf_jobs("uv", {
"script": """
# /// script
# dependencies = ["trl>=0.12.0", "peft>=0.7.0", "trackio"]
# ///
from datasets import load_dataset
from peft import LoraConfig
from trl import SFTTrainer, SFTConfig
import trackio
dataset = load_dataset("trl-lib/Capybara", split="train")
# Create train/eval split for monitoring
dataset_split = dataset.train_test_split(test_size=0.1, seed=42)
trainer = SFTTrainer(
model="Qwen/Qwen2.5-0.5B",
train_dataset=dataset_split["train"],
eval_dataset=dataset_split["test"],
peft_config=LoraConfig(r=16, lora_alpha=32),
args=SFTConfig(
output_dir="my-model",
push_to_hub=True,
hub_model_id="username/my-model",
num_train_epochs=3,
eval_strategy="steps",
eval_steps=50,
report_to="trackio",
project="meaningful_prject_name", # project name for the training name (trackio)
run_name="meaningful_run_name", # descriptive name for the specific training run (trackio)
)
)
trainer.train()
trainer.push_to_hub()
""",
"flavor": "a10g-large",
"timeout": "2h",
"secrets": {"HF_TOKEN": "$HF_TOKEN"}
})
Benefits: Direct MCP tool usage, clean code, dependencies declared inline (PEP 723), no file saving required, full control
When to use: Default choice for all training tasks in Claude Code, custom training logic, any scenario requiring hf_jobs()
Working with Scripts
⚠️ Important: The script parameter accepts either inline code (as shown above) OR a URL. Local file paths do NOT work.
Why local paths don't work:
Jobs run in isolated Docker containers without access to your local filesystem. Scripts must be:
Inline code (recommended for custom training)
Publicly accessible URLs
Private repo URLs (with HF_TOKEN)
Common mistakes:
# ❌ These will all fail
hf_jobs("uv", {"script": "train.py"})
hf_jobs("uv", {"script": "./scripts/train.py"})
hf_jobs("uv", {"script": "/path/to/train.py"})
Correct approaches:
# ✅ Inline code (recommended)
hf_jobs("uv", {"script": "# /// script\n# dependencies = [...]\n# ///\n\n<your code>"})
# ✅ From Hugging Face Hub
hf_jobs("uv", {"script": "https://huggingface.co/user/repo/resolve/main/train.py"})
# ✅ From GitHub
hf_jobs("uv", {"script": "https://raw.githubusercontent.com/user/repo/main/train.py"})
# ✅ From Gist
hf_jobs("uv", {"script": "https://gist.githubusercontent.com/user/id/raw/train.py"})
Benefits: No code to write, maintained by TRL team, production-tested
When to use: Standard TRL training, quick experiments, don't need custom code
Available: Scripts are available from https://github.com/huggingface/trl/tree/main/examples/scripts
Finding More UV Scripts on Hub
The uv-scripts organization provides ready-to-use UV scripts stored as datasets on Hugging Face Hub:
# Discover available UV script collections
dataset_search({"author": "uv-scripts", "sort": "downloads", "limit": 20})
# Explore a specific collection
hub_repo_details(["uv-scripts/classification"], repo_type="dataset", include_readme=True)
Popular collections: ocr, classification, synthetic-data, vllm, dataset-creation
Approach 3: HF Jobs CLI (Direct Terminal Commands)
When the hf_jobs() MCP tool is unavailable, use the hf jobs CLI directly.
⚠️ CRITICAL: CLI Syntax Rules
# ✅ CORRECT syntax - flags BEFORE script URL
hf jobs uv run --flavor a10g-large --timeout 2h --secrets HF_TOKEN "https://example.com/train.py"# ❌ WRONG - "run uv" instead of "uv run"
hf jobs run uv "https://example.com/train.py" --flavor a10g-large
# ❌ WRONG - flags AFTER script URL (will be ignored!)
hf jobs uv run "https://example.com/train.py" --flavor a10g-large
# ❌ WRONG - "--secret" instead of "--secrets" (plural)
hf jobs uv run --secret HF_TOKEN "https://example.com/train.py"
Key syntax rules:
Command order is hf jobs uv run (NOT hf jobs run uv)
All flags (--flavor, --timeout, --secrets) must come BEFORE the script URL
Use --secrets (plural), not --secret
Script URL must be the last positional argument
Complete CLI example:
hf jobs uv run \
--flavor a10g-large \
--timeout 2h \
--secrets HF_TOKEN \
"https://huggingface.co/user/repo/resolve/main/train.py"
Check job status via CLI:
hf jobs ps # List all jobs
hf jobs logs <job-id> # View logs
hf jobs inspect <job-id> # Job details
hf jobs cancel <job-id> # Cancel a job
Approach 4: TRL Jobs Package (Simplified Training)
The trl-jobs package provides optimized defaults and one-liner training.
Benefits: Pre-configured settings, automatic Trackio integration, automatic Hub push, one-line commands
When to use: User working in terminal directly (not Claude Code context), quick local experimentation
Repository:https://github.com/huggingface/trl-jobs
⚠️ In Claude Code context, prefer using hf_jobs() MCP tool (Approach 1) when available.
Output includes estimated time, cost, recommended timeout (with buffer), and optimization suggestions.
When to offer: User planning a job, asks about cost/time, choosing hardware, job will run >1 hour or cost >$5
Example Training Scripts
Production-ready templates with all best practices:
Load these scripts for correctly:
scripts/train_sft_example.py - Complete SFT training with Trackio, LoRA, checkpoints
scripts/train_dpo_example.py - DPO training for preference learning
scripts/train_grpo_example.py - GRPO training for online RL
These scripts demonstrate proper Hub saving, Trackio integration, checkpoint management, and optimized parameters. Pass their content inline to hf_jobs() or use as templates for custom scripts.
Monitoring and Tracking
Trackio provides real-time metrics visualization. See references/trackio_guide.md for complete setup guide.
Key points:
Add trackio to dependencies
Configure trainer with report_to="trackio" and run_name="meaningful_name"
Trackio Configuration Defaults
Use sensible defaults unless user specifies otherwise. When generating training scripts with Trackio:
Default Configuration:
Space ID: {username}/trackio (use "trackio" as default space name)
Run naming: Unless otherwise specified, name the run in a way the user will recognize (e.g., descriptive of the task, model, or purpose)
Config: Keep minimal - only include hyperparameters and model/dataset info
Project Name: Use a Project Name to associate runs with a particular Project
User overrides: If user requests specific trackio configuration (custom space, run naming, grouping, or additional config), apply their preferences instead of defaults.
This is useful for managing multiple jobs with the same configuration or keeping training scripts portable.
See references/trackio_guide.md for complete documentation including grouping runs for experiments.
Check Job Status
# List all jobs
hf_jobs("ps")
# Inspect specific job
hf_jobs("inspect", {"job_id": "your-job-id"})
# View logs
hf_jobs("logs", {"job_id": "your-job-id"})
Remember: Wait for user to request status checks. Avoid polling repeatedly.
Dataset Validation
Validate dataset format BEFORE launching GPU training to prevent the #1 cause of training failures: format mismatches.
Why Validate
50%+ of training failures are due to dataset format issues
DPO especially strict: requires exact column names (prompt, chosen, rejected)
Failed GPU jobs waste $1-10 and 30-60 minutes
Validation on CPU costs ~$0.01 and takes <1 minute
When to Validate
ALWAYS validate for:
Unknown or custom datasets
DPO training (CRITICAL - 90% of datasets need mapping)
Any dataset not explicitly TRL-compatible
Skip validation for known TRL datasets:
trl-lib/ultrachat_200k, trl-lib/Capybara, HuggingFaceH4/ultrachat_200k, etc.
See references/training_patterns.md for detailed examples including:
Quick demo (5-10 minutes)
Production with checkpoints
Multi-GPU training
DPO training (preference learning)
GRPO training (online RL)
Common Failure Modes
Out of Memory (OOM)
Fix (try in order):
Reduce batch size: per_device_train_batch_size=1, increase gradient_accumulation_steps=8. Effective batch size is per_device_train_batch_size x gradient_accumulation_steps. For best performance keep effective batch size close to 128.
Enable: gradient_checkpointing=True
Upgrade hardware: t4-small → l4x1, a10g-small → a10g-large etc.
Dataset Misformatted
Fix:
Validate first with dataset inspector:
uv run https://huggingface.co/datasets/mcp-tools/skills/raw/main/dataset_inspector.py \
--dataset name --split train
Submit scripts inline - The script parameter accepts Python code directly; no file saving required unless user requests
Jobs are asynchronous - Don't wait/poll; let user check when ready
Always set timeout - Default 30 min is insufficient; minimum 1-2 hours recommended
Always enable Hub push - Environment is ephemeral; without push, all results lost
Include Trackio - Use example scripts as templates for real-time monitoring
Offer cost estimation - When parameters are known, use scripts/estimate_cost.py
Use UV scripts (Approach 1) - Default to hf_jobs("uv", {...}) with inline scripts; TRL maintained scripts for standard training; avoid bash trl-jobs commands in Claude Code
Use hf_doc_fetch/hf_doc_search for latest TRL documentation
Validate dataset format before training with dataset inspector (see Dataset Validation section)
Choose appropriate hardware for model size; use LoRA for models >7B
