Hugging Face Model Trainer

TRL Training on Hugging Face Jobs

Train language models using TRL on managed Hugging Face infrastructure. Models train on cloud GPUs and results save to Hugging Face Hub automatically.

Key Rules

1. Submit via hf_jobs() MCP tool — use hf_jobs("uv", {...}) with inline Python code in the script parameter. Local file paths do not work (jobs run in isolated Docker containers). When a user asks to train a model, create the script and submit immediately.
2. Include Trackio in every training script for real-time monitoring. Use example scripts in scripts/ as templates.
3. Enable Hub push — the training environment is ephemeral; unpushed results are lost. Set push_to_hub=True, hub_model_id, and secrets={"HF_TOKEN": "$HF_TOKEN"}.
4. Set adequate timeout — default 30 min is too short for most training. Add 20-30% buffer to estimated time.
5. Validate unknown datasets before GPU training to prevent format failures. See references/dataset_validation.md.
6. Report after submission — provide job ID, monitoring URL, estimated time. Wait for user to request status checks.

Training Methods

For method details: references/training_methods.md or hf_doc_search("query", product="trl")

When to Use Unsloth

Use Unsloth when VRAM is limited, speed matters, training large models (>13B), or training VLMs. ~2x faster, ~60% less VRAM. See references/unsloth.md.

Prerequisites

Account & Auth

• HF account with Pro, Team, or Enterprise plan
• Authenticated: verify with hf_whoami()
• Token with write permissions; pass secrets={"HF_TOKEN": "$HF_TOKEN"} in job config

Dataset

• Dataset on Hub or loadable via datasets.load_dataset()
• Format matching training method (see table above)
• Validate unknown datasets before GPU training (see references/dataset_validation.md)
• Size appropriate for hardware (demo: 50-100 examples on t4-small; production: 1K-10K+ on a10g/a100)

Job Submission: UV Scripts (Primary Approach)

UV scripts use PEP 723 inline dependencies for self-contained training:

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")
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_project_name",
        run_name="meaningful_run_name",
    )
)

trainer.train()
trainer.push_to_hub()
""",
    "flavor": "a10g-large",
    "timeout": "2h",
    "secrets": {"HF_TOKEN": "$HF_TOKEN"}
})

Demo tip: For quick demos on smaller GPUs (t4-small), omit eval_dataset and eval_strategy to save ~40% memory.

Sequence Length

TRL config classes use max_length (not max_seq_length):

SFTConfig(max_length=512)   # ✅ Truncate to 512 tokens
SFTConfig(max_seq_length=512)  # ❌ TypeError

Default max_length=1024 works well for most training. Override for longer context, memory constraints, or vision models (max_length=None).

Alternative Approaches

See references/cli_usage.md for:

• TRL maintained scripts — battle-tested official examples from URLs
• HF Jobs CLI — hf jobs uv run terminal commands (when MCP unavailable)
• TRL Jobs package — trl-jobs sft one-liner training
• UV Scripts on Hub — ready-to-use scripts from uv-scripts organization

Hardware Selection

Use LoRA/PEFT for models >7B. Multi-GPU is handled automatically by TRL/Accelerate.

All flavors: cpu-basic/upgrade/performance/xl, t4-small/medium, l4x1/x4, a10g-small/large/largex2/largex4, a100-large, h100/h100x8

See references/hardware_guide.md for detailed specifications.

Hub Saving

The training environment is ephemeral — all files are deleted when the job ends.

In training config:

SFTConfig(
    push_to_hub=True,
    hub_model_id="username/model-name",
    hub_strategy="every_save",  # optional: push checkpoints
)

In job submission:

{"secrets": {"HF_TOKEN": "$HF_TOKEN"}}

Checklist: push_to_hub=True ✓ | hub_model_id set ✓ | secrets has HF_TOKEN ✓ | write access ✓

See references/hub_saving.md for troubleshooting.

Timeout Guidelines

On timeout, the job is killed immediately and unsaved progress is lost. Add 20-30% buffer.

{"timeout": "2h"}  # formats: "90m", "2h", "1.5h", or seconds as integer

Cost Estimation

Offer to estimate cost when planning jobs with known parameters:

uv run scripts/estimate_cost.py \
  --model meta-llama/Llama-2-7b-hf \
  --dataset trl-lib/Capybara \
  --hardware a10g-large \
  --dataset-size 16000 \
  --epochs 3

Output: estimated time, cost, recommended timeout, optimization suggestions.

Monitoring

Include Trackio in every script (report_to="trackio"). Default config:

• Space ID: {username}/trackio
• Run naming: descriptive of task, model, or purpose
• Project name: group related runs

See references/trackio_guide.md for complete setup and experiment grouping.

Job Status

hf_jobs("ps")                                    # List all jobs
hf_jobs("inspect", {"job_id": "your-job-id"})    # Job details
hf_jobs("logs", {"job_id": "your-job-id"})       # View logs

Dataset Validation

Validate format before GPU training — the #1 cause of training failures:

hf_jobs("uv", {
    "script": "https://huggingface.co/datasets/mcp-tools/skills/raw/main/dataset_inspector.py",
    "script_args": ["--dataset", "username/dataset-name", "--split", "train"]
})

Output markers: ✓ READY (use directly) | ✗ NEEDS MAPPING (code provided) | ✗ INCOMPATIBLE

Skip validation for known TRL datasets (trl-lib/Capybara, trl-lib/ultrachat_200k, etc.).

See references/dataset_validation.md for full workflow and DPO mapping examples.

GGUF Conversion

Convert trained models for local inference (Ollama, LM Studio, llama.cpp):

hf_jobs("uv", {
    "script": "<see references/gguf_conversion.md for complete script>",
    "flavor": "a10g-large",
    "timeout": "45m",
    "secrets": {"HF_TOKEN": "$HF_TOKEN"},
    "env": {
        "ADAPTER_MODEL": "username/my-finetuned-model",
        "BASE_MODEL": "Qwen/Qwen2.5-0.5B",
        "OUTPUT_REPO": "username/my-model-gguf"
    }
})

See references/gguf_conversion.md for quantization options, hardware requirements, and troubleshooting.

Common Failure Modes

See references/troubleshooting.md for detailed solutions.

Local Script Dependencies

pip install -r requirements.txt

References

In This Skill

• references/training_methods.md — SFT, DPO, GRPO, KTO, PPO, Reward Modeling overview
• references/training_patterns.md — Common training patterns and examples
• references/dataset_validation.md — Dataset validation workflow
• references/cli_usage.md — CLI, TRL scripts, alternative submission methods
• references/unsloth.md — Unsloth for fast training (~2x speed, 60% less VRAM)
• references/gguf_conversion.md — GGUF conversion guide
• references/trackio_guide.md — Trackio monitoring setup
• references/hardware_guide.md — Hardware specs and selection
• references/hub_saving.md — Hub auth troubleshooting
• references/troubleshooting.md — Common issues and solutions

Scripts

• scripts/train_sft_example.py — Production SFT template
• scripts/train_dpo_example.py — Production DPO template
• scripts/train_grpo_example.py — Production GRPO template
• scripts/unsloth_sft_example.py — Unsloth training template
• scripts/estimate_cost.py — Time and cost estimator
• scripts/convert_to_gguf.py — GGUF conversion script

External

• Dataset Inspector — validate dataset format
• TRL Documentation
• TRL Jobs Training Guide
• TRL Jobs Package
• HF Jobs Documentation
• TRL Example Scripts
• UV Scripts Guide
• UV Scripts Organization