huggingface-llm-trainer

Name: huggingface-llm-trainer
Author: huggingface

1.4kinstalls722forks

How to install huggingface-llm-trainer?

huggingface/skills/huggingface-llm-trainer

$npx -y skills add huggingface/skills --skill huggingface-llm-trainer

Installs into the current project.

Summary

Train or fine-tune language and vision models using TRL (Transformer Reinforcement Learning) or Unsloth with Hugging Face Jobs infrastructure. Covers SFT, DPO, GRPO and reward modeling training methods, plus GGUF conversion for local deployment. Includes guidance on the TRL Jobs

Files · 1

View on GitHub

SKILL.md

1	# TRL Training on Hugging Face Jobs
2
3	## Overview
4
5	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.
6
7	TRL provides multiple training methods:
8	- SFT (Supervised Fine-Tuning) - Standard instruction tuning
9	- DPO (Direct Preference Optimization) - Alignment from preference data
10	- GRPO (Group Relative Policy Optimization) - Online RL training
11	- Reward Modeling - Train reward models for RLHF
12
13	For detailed TRL method documentation:
14	```python
15	hf_doc_search("your query", product="trl")
16	hf_doc_fetch("https://huggingface.co/docs/trl/sft_trainer") # SFT
17	hf_doc_fetch("https://huggingface.co/docs/trl/dpo_trainer") # DPO
18	# etc.
19	```
20
21	See also: `references/training_methods.md` for method overviews and selection guidance
22
23	## When to Use This Skill
24
25	Use this skill when users want to:
26	- Fine-tune language models on cloud GPUs without local infrastructure
27	- Train with TRL methods (SFT, DPO, GRPO, etc.)
28	- Run training jobs on Hugging Face Jobs infrastructure
29	- Convert trained models to GGUF for local deployment (Ollama, LM Studio, llama.cpp)
30	- Ensure trained models are permanently saved to the Hub
31	- Use modern workflows with optimized defaults
32
33	### When to Use Unsloth
34
35	Use Unsloth (`references/unsloth.md`) instead of standard TRL when:
36	- Limited GPU memory - Unsloth uses ~60% less VRAM
37	- Speed matters - Unsloth is ~2x faster
38	- Training large models (>13B) - memory efficiency is critical
39	- Training Vision-Language Models (VLMs) - Unsloth has `FastVisionModel` support
40
41	See `references/unsloth.md` for complete Unsloth documentation and `scripts/unsloth_sft_example.py` for a production-ready training script.
42
43	## Key Directives
44
45	When assisting with training jobs:
46
47	1. 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()`.
48
49	2. Always include Trackio - Every training script should include Trackio for real-time monitoring. Use example scripts in `scripts/` as templates.
50
51	3. Provide job details after submission - After submitting, provide job ID, monitoring URL, estimated time, and note that the user can request status checks later.
52
53	4. Use example scripts as templates - Reference `scripts/train_sft_example.py`, `scripts/train_dpo_example.py`, etc. as starting points.
54
55	## Local Script Execution
56
57	Repository scripts use PEP 723 inline dependencies. Run them with `uv run`:
58	```bash
59	uv run scripts/estimate_cost.py --help
60	uv run scripts/dataset_inspector.py --help
61	```
62
63	## Prerequisites Checklist
64
65	Before starting any training job, verify:
66
67	### ✅ Account & Authentication
68	- Hugging Face Account with [Pro](https://hf.co/pro), [Team](https://hf.co/enterprise), or [Enterprise](https://hf.co/enterprise) plan (Jobs require paid plan)
69	- Authenticated login: Check with `hf_whoami()`
70	- HF_TOKEN for Hub Push ⚠️ CRITICAL - Training environment is ephemeral, must push to Hub or ALL training results are lost
71	- Token must have write permissions
72	- MUST pass `secrets={"HF_TOKEN": "$HF_TOKEN"}` in job config to make token available (the `$HF_TOKEN` syntax
73	references your actual token value)
74
75	### ✅ Dataset Requirements
76	- Dataset must exist on Hub or be loadable via `datasets.load_dataset()`
77	- Format must match training method (SFT: "messages"/text/prompt-completion; DPO: chosen/rejected; GRPO: prompt-only)
78	- ALWAYS validate unknown datasets before GPU training to prevent format failures (see Dataset Validation section below)
79	- Size appropriate for hardware (Demo: 50-100 examples on t4-small; Production: 1K-10K+ on a10g-large/a100-large)
80
81	### ⚠️ Critical Settings
82	- 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 t

People who liked this also viewed

6 picks

Type

huggingface-llm-trainer

1.4kinstalls722forks

View on GitHub

How to install huggingface-llm-trainer?

huggingface/skills/huggingface-llm-trainer

$npx -y skills add huggingface/skills --skill huggingface-llm-trainer

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# TRL Training on Hugging Face Jobs
2
3	## Overview
4
5	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.
6
7	TRL provides multiple training methods:
8	- SFT (Supervised Fine-Tuning) - Standard instruction tuning
9	- DPO (Direct Preference Optimization) - Alignment from preference data
10	- GRPO (Group Relative Policy Optimization) - Online RL training
11	- Reward Modeling - Train reward models for RLHF
12
13	For detailed TRL method documentation:
14	```python
15	hf_doc_search("your query", product="trl")
16	hf_doc_fetch("https://huggingface.co/docs/trl/sft_trainer") # SFT
17	hf_doc_fetch("https://huggingface.co/docs/trl/dpo_trainer") # DPO
18	# etc.
19	```
20
21	See also: `references/training_methods.md` for method overviews and selection guidance
22
23	## When to Use This Skill
24
25	Use this skill when users want to:
26	- Fine-tune language models on cloud GPUs without local infrastructure
27	- Train with TRL methods (SFT, DPO, GRPO, etc.)
28	- Run training jobs on Hugging Face Jobs infrastructure
29	- Convert trained models to GGUF for local deployment (Ollama, LM Studio, llama.cpp)
30	- Ensure trained models are permanently saved to the Hub
31	- Use modern workflows with optimized defaults
32
33	### When to Use Unsloth
34
35	Use Unsloth (`references/unsloth.md`) instead of standard TRL when:
36	- Limited GPU memory - Unsloth uses ~60% less VRAM
37	- Speed matters - Unsloth is ~2x faster
38	- Training large models (>13B) - memory efficiency is critical
39	- Training Vision-Language Models (VLMs) - Unsloth has `FastVisionModel` support
40
41	See `references/unsloth.md` for complete Unsloth documentation and `scripts/unsloth_sft_example.py` for a production-ready training script.
42
43	## Key Directives
44
45	When assisting with training jobs:
46
47	1. 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()`.
48
49	2. Always include Trackio - Every training script should include Trackio for real-time monitoring. Use example scripts in `scripts/` as templates.
50
51	3. Provide job details after submission - After submitting, provide job ID, monitoring URL, estimated time, and note that the user can request status checks later.
52
53	4. Use example scripts as templates - Reference `scripts/train_sft_example.py`, `scripts/train_dpo_example.py`, etc. as starting points.
54
55	## Local Script Execution
56
57	Repository scripts use PEP 723 inline dependencies. Run them with `uv run`:
58	```bash
59	uv run scripts/estimate_cost.py --help
60	uv run scripts/dataset_inspector.py --help
61	```
62
63	## Prerequisites Checklist
64
65	Before starting any training job, verify:
66
67	### ✅ Account & Authentication
68	- Hugging Face Account with [Pro](https://hf.co/pro), [Team](https://hf.co/enterprise), or [Enterprise](https://hf.co/enterprise) plan (Jobs require paid plan)
69	- Authenticated login: Check with `hf_whoami()`
70	- HF_TOKEN for Hub Push ⚠️ CRITICAL - Training environment is ephemeral, must push to Hub or ALL training results are lost
71	- Token must have write permissions
72	- MUST pass `secrets={"HF_TOKEN": "$HF_TOKEN"}` in job config to make token available (the `$HF_TOKEN` syntax
73	references your actual token value)
74
75	### ✅ Dataset Requirements
76	- Dataset must exist on Hub or be loadable via `datasets.load_dataset()`
77	- Format must match training method (SFT: "messages"/text/prompt-completion; DPO: chosen/rejected; GRPO: prompt-only)
78	- ALWAYS validate unknown datasets before GPU training to prevent format failures (see Dataset Validation section below)
79	- Size appropriate for hardware (Demo: 50-100 examples on t4-small; Production: 1K-10K+ on a10g-large/a100-large)
80
81	### ⚠️ Critical Settings
82	- 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 t

People who liked this also viewed

6 picks

Type