New: 97,000+ skills indexed from GitHub repos with 100+ stars · every skill safety-checkedBrowse top skillsTop skills
Esc to closeBrowse all skills
Submit a skillBrowse skills
Submit a skillBrowse skills

Agent skill

Supervised Preference Training

by VectorSpaceLab in VectorSpaceLab/AREX-Skill

Build and review OpenRLHF supervised/preference training plans for SFT, reward models, DPO, IPO, and cDPO.

Apache-2.0Auto-check passedAI & LLM Engineering

Install Supervised Preference Training

skills CLI
$ npx skills add VectorSpaceLab/AREX-Skill --skill supervised-preference-training -a claude-code

Project install by default; add -g for ~/.claude/skills/.

GitHub CLI
$ gh skill install VectorSpaceLab/AREX-Skill supervised-preference-training --agent claude-code

Project scope by default; add --scope user for a personal install. Needs GitHub CLI 2.90.0 or later (public preview).

Manual copy
$ git clone --depth 1 https://github.com/VectorSpaceLab/AREX-Skill.git skills-src && mkdir -p .claude/skills && cp -r skills-src/skills/repositories/repo-skills/openrlhf/sub-skills/supervised-preference-training .claude/skills/supervised-preference-training && rm -rf skills-src

Use ~/.claude/skills/ instead of .claude/skills for a personal install. The folder must contain SKILL.md.

Claude Code skills documentation · loads skills from .claude/skills/

Facts

Skill name
supervised-preference-training
GitHub stars
328
Token cost
~1k tokens
SKILL.md length
366 words
Files
5 (incl. scripts, references)
Skills in repo
159
Repo updated
First seen
Licence
Apache-2.0

At a glance

Build and review OpenRLHF supervised/preference training plans for SFT, reward models, DPO, IPO, and cDPO.

  • Trainsft/trainrm/traindpo CLI construction
  • SKILL.md covers Route First, Core References, Command Builder and Safety Checklist, plus 1 more section
  • Runs Python scripts from its folder; calls python
  • LoRA/packing/DeepSpeed/checkpoint/logging choices

What it does

Supervised Preference Training is an agent skill from VectorSpaceLab/AREX-Skill. Build and review OpenRLHF supervised/preference training plans for SFT, reward models, DPO, IPO, and cDPO. Use for trainsft/trainrm/traindpo CLI construction, LoRA/packing/DeepSpeed/checkpoint/logging choices, and preflight checks before expensive GPU training.

Its SKILL.md is about 1k tokens, which your agent loads only when the skill is triggered. The skill folder holds 6 other files, including scripts and reference files (for example `references/cli-reference.md`, `references/training-workflows.md` and `references/troubleshooting.md`).

It sits in AI & LLM Engineering, covering Fine-tuning, Reinforcement learning and Deep learning. The repository describes itself as: A Skill Library for Automated Machine Learning. The licence is Apache-2.0.

When your agent uses it

  • Trainsft/trainrm/traindpo CLI construction
  • LoRA/packing/DeepSpeed/checkpoint/logging choices
  • Preflight checks before expensive GPU training

Example prompts

  • “/supervised-preference-training”

Requirements

  • Python 3
  • Docker

What it can do on your machine

Read from SKILL.md and the folder at commit ac3fe1a. It shows what the files ask for, not the result of running them.

  • Tool permissions

    Pre-approves nothing: there is no allowed-tools line, so your agent's usual permission prompts apply.

    From allowed-tools in the SKILL.md frontmatter.

  • Runs code

    Ships 1 file in scripts/ (Python), which the agent can run.

    Shell commands in SKILL.md call:

    • python

    From the folder's file list and the shell code blocks in SKILL.md.

  • Network

    No URLs in SKILL.md.

    From URLs in SKILL.md, links to its own repository left out.

  • Credentials

    Names no API keys, tokens, secrets or passwords.

    From names ending in _API_KEY, _TOKEN, _SECRET, _KEY or _PASSWORD in SKILL.md.

Context cost

Supervised Preference Training loads about 1k tokens when it runs, and up to ~6.2k if it reads all its reference files. Until then it costs about 74 tokens; SKILL.md has 366 words of instructions outside code blocks.

Always · name and description, kept in context so the agent knows when to use it
~74
When it runs · the whole SKILL.md, loaded when a task matches
~1k
With references · SKILL.md plus every file in references/, read only if the agent opens them
~6.2k

Estimates: characters ÷ 4, the usual rule of thumb; real counts depend on the model's tokenizer. Scripts and assets cost tokens only if the agent reads them.

Safety

Auto-check passed

The automated check found no risky patterns in SKILL.md.

Automated static check — not a guarantee. Review scripts before installing. It scans the text of SKILL.md for risky patterns (piping downloads into a shell, reading credential files, hidden Unicode, destructive commands); the scripts in this folder are not scanned.

SKILL.md

The full file from VectorSpaceLab/AREX-Skill at commit ac3fe1a, republished under its Apache-2.0 licence (© VectorSpaceLab). 366 words, ~1,032 tokens.

Download SKILL.mdSave it as .claude/skills/supervised-preference-training/SKILL.md (or your agent's skills folder). This skill also uses 4 other files; get the full folder from GitHub.
name
supervised-preference-training
description
Build and review OpenRLHF supervised/preference training plans for SFT, reward models, DPO, IPO, and cDPO. Use for train_sft/train_rm/train_dpo CLI construction, LoRA/packing/DeepSpeed/checkpoint/logging choices, and preflight checks before expensive GPU training.
disable-model-invocation
true
metadata.disco-role
operating
license
Apache 2.0

Supervised Preference Training

Use this sub-skill when the user is preparing OpenRLHF SFT, reward-model, DPO, IPO, or cDPO training and needs command construction, flag review, or risk checks. Treat all actual training launches as expensive GPU/network actions.

Route First

  • For SFT, use openrlhf.cli.train_sft with prompt/completion keys such as --data.input_key and --data.output_key.
  • For reward-model training, use openrlhf.cli.train_rm with preference keys such as --data.chosen_key and --data.rejected_key.
  • For DPO, IPO, and cDPO, use openrlhf.cli.train_dpo; add --model.ipo_enable for IPO and --model.label_smoothing for cDPO.
  • For detailed dataset schema conversion, chat templates, multiturn examples, and key mapping, route to the data-preparation sub-skill.
  • For PPO, REINFORCE++, Ray, vLLM, remote actors, and agent training, route to the rl-agent-training sub-skill.
  • For installation, FlashAttention/Liger/RingAttention dependencies, Ray clusters, serving, Docker, and environment repair, route to operations-and-utilities.

Core References

  • Read references/training-workflows.md for SFT/RM/DPO workflow recipes, source-backed shell patterns, and preflight order.
  • Read references/cli-reference.md for current CLI flag names and source-backed defaults.
  • Read references/troubleshooting.md for common failure modes before recommending a training run.
  • Use scripts/build_training_command.py to print a safe command skeleton without importing OpenRLHF or starting training.

Command Builder

The bundled helper is safe for planning and help-only validation:

bash
python skills/openrlhf/sub-skills/supervised-preference-training/scripts/build_training_command.py sft --model MODEL --dataset DATASET --output-dir OUT
python skills/openrlhf/sub-skills/supervised-preference-training/scripts/build_training_command.py rm --model MODEL --dataset DATASET --output-dir OUT
python skills/openrlhf/sub-skills/supervised-preference-training/scripts/build_training_command.py dpo --model MODEL --dataset DATASET --output-dir OUT --ref-model REF --beta 0.1 --label-smoothing 0.1 --nll-loss-coef 0.05

It prints deepspeed --module openrlhf.cli.train_* ... commands for review. It does not check GPU availability, download models, import OpenRLHF, or execute the result.

Show full SKILL.md (158 more words)Show less

Safety Checklist

Before approving or launching a generated command:

  • Confirm model and dataset identifiers/paths are intended and access-controlled; Hugging Face or ModelScope names may trigger network downloads.
  • Confirm dataset keys match the selected trainer; SFT uses input/output keys, RM/DPO use chosen/rejected preference keys.
  • Confirm --train.batch_size is global and --train.micro_batch_size is per GPU; reduce micro-batch size or use ZeRO-3/offload/LoRA for OOM.
  • Confirm optional kernels (flash_attention_2, Liger, RingAttention) and 4-bit/LoRA dependencies exist before using their flags.
  • Confirm checkpoint behavior: --ckpt.save_steps -1 disables periodic DeepSpeed checkpoint saves; --ckpt.save_hf writes HF-format saves at checkpoint intervals.
  • Prefer source-backed --model.model_name_or_path over older README snippets that may mention --actor.model_name_or_path for these CLIs.

Evidence Base

This sub-skill is based on OpenRLHF training entrypoints openrlhf.cli.train_sft, openrlhf.cli.train_rm, and openrlhf.cli.train_dpo; trainers sft_trainer.py, rm_trainer.py, and dpo_trainer.py; README SFT/RM/DPO examples; and example shell recipes for SFT, RM, DPO, and SFT LoRA. The installed package import was verified for openrlhf version 0.10.4, but full dependency and GPU runtime readiness were not verified.

© VectorSpaceLab, Apache-2.0. Rendered from Markdown: HTML in the file is shown as text, images as links, and headings moved down two levels. Raw file

Files

SKILL.md and 4 other files (scripts, references) in skills/repositories/repo-skills/openrlhf/sub-skills/supervised-preference-training of VectorSpaceLab/AREX-Skill.

  • SKILL.md
  • references/cli-reference.md
  • references/training-workflows.md
  • references/troubleshooting.md
  • scripts/build_training_command.py

Open the folder on GitHubat commit ac3fe1a

Compare with similar skills

Supervised Preference Training next to the 5 skills that share the most tags, products or categories with it. Stars are the repository's; “used in” counts other GitHub owners with a copy.

Supervised Preference Training compared with similar skills
SkillStarsUsed inTokensAuto-checkLicenceRepo updated
Supervised Preference Training this skillVectorSpaceLab/AREX-Skill328—~1kAutomated safety check: PassApache-2.0
Grpobenchflow-ai/skillsbench1.8k—~1.1kAutomated safety check: PassApache-2.0
Slime Useryzlnew/infra-skills149—~3.2kAutomated safety check: PassNone
Hugging Face LLM Trainerhuggingface/skills11k3 repos~7.2kAutomated safety check: PassApache-2.0
Train RlOpenPipe/ART11k—~2.4kAutomated safety check: PassApache-2.0
Qwopus27b Rl TrainingR6410418/Jackrong-llm-finetuning-guide1.7k—~830Automated safety check: PassApache-2.0

Similar skills

  • Grpo

    benchflow-ai/skillsbench

    Reference for the GRPO (Group Relative Policy Optimization) algorithm.

    1.8k GitHub stars~1.1k tokensUpdated 2 mo ago
    AI & LLM EngineeringAuto-check passed

  • Slime User

    yzlnew/infra-skills

    Guide for using SLIME (LLM post-training framework for RL Scaling).

    149 GitHub stars~3.2k tokensUpdated 3 mo ago
    AI & LLM EngineeringAuto-check passed

  • Hugging Face LLM Trainer

    huggingface/skills

    Official

    Trains or fine-tunes language and vision models with TRL or Unsloth on Hugging Face Jobs cloud GPUs, then converts the results to GGUF.

    11k GitHub starsUsed in 3 repos~7.2k tokens
    AI & LLM EngineeringAuto-check passed

  • Train Rl

    OpenPipe/ART

    RL training reference for the ART framework. An agent skill from OpenPipe/ART.

    11k GitHub stars~2.4k tokensUpdated 2 days ago
    AI & LLM EngineeringAuto-check passed

  • Qwopus27b Rl Training

    R6410418/Jackrong-llm-finetuning-guide

    Prepare, validate, launch-plan, monitor, resume, and stop configurable Qwopus 27B reinforcement-learning workflows for GRPO or GSPO.

    1.7k GitHub stars~830 tokensUpdated 2 mo ago
    AI & LLM EngineeringAuto-check passed

  • Fine Tuning With Trl

    Orchestra-Research/AI-Research-SKILLs

    Fine-tune LLMs using reinforcement learning with TRL - SFT for instruction tuning, DPO for preference alignment, PPO/GRPO for reward optimization, and reward model training.

    13k GitHub starsUsed in 7 repos~2.9k tokens
    AI & LLM EngineeringAuto-check passed

More from VectorSpaceLab/AREX-Skill

All 159 skills in this repo

  • Agent Lightning

    VectorSpaceLab/AREX-Skill

    Use this repo skill for Agent Lightning package tasks: authoring trainable agents, tracing rewards and spans, running LightningStore/Trainer loops, using agl CLI services, choosing examples, and…

    328 GitHub stars~1.4k tokensUpdated 1 mo ago
    Auto-check passed

  • Agent Tools

    VectorSpaceLab/AREX-Skill

    A skill your agent uses when configuring LiteLLM for MCP tools, A2A agents, Claude Code/Cursor agent gateway traffic, MCP auth/OAuth, tool permissions, semantic filtering, or agent-specific proxy…

    328 GitHub stars~1.2k tokensUpdated 1 mo ago
    Auto-check passed

  • Agents And Awel

    VectorSpaceLab/AREX-Skill

    Build and debug DB-GPT agents, tools, skills, teams, and AWEL workflows, including deterministic local DAG runs and HTTP-trigger topology without assuming an LLM, credential, or external service.

    328 GitHub stars~2.3k tokensUpdated 1 mo ago
    Auto-check passed

  • Agents And Middleware

    VectorSpaceLab/AREX-Skill

    Work on the actively maintained LangChain v1 agent package: initchatmodel, createagent, structured output, tools, middleware, embeddings initialization, provider routing, and agent runtime…

    328 GitHub stars~1.2k tokensUpdated 1 mo ago
    Auto-check passed

  • Agents Workflows

    VectorSpaceLab/AREX-Skill

    A skill your agent uses for giskard.agents async chat workflows, tools, prompt templates, structured outputs, retries, rate limiting, embeddings, and optional LiteLLM backend.

    328 GitHub stars~500 tokensUpdated 1 mo ago
    Auto-check passed

  • Alphafold3

    VectorSpaceLab/AREX-Skill

    A skill your agent uses for AlphaFold 3 input preparation, prediction command planning, output interpretation, and Python API inspection.

    328 GitHub stars~1.2k tokensUpdated 1 mo ago
    Auto-check passed

Categories

Questions about Supervised Preference Training

What does Supervised Preference Training do?

Build and review OpenRLHF supervised/preference training plans for SFT, reward models, DPO, IPO, and cDPO. Supervised Preference Training is an agent skill from VectorSpaceLab/AREX-Skill. Build and review OpenRLHF supervised/preference training plans for SFT, reward models, DPO, IPO, and cDPO.

When should I use Supervised Preference Training?

Supervised Preference Training fits situations like: trainsft/trainrm/traindpo CLI construction; loRA/packing/DeepSpeed/checkpoint/logging choices; preflight checks before expensive GPU training.

How do I install Supervised Preference Training in Claude Code?

Run `npx skills add VectorSpaceLab/AREX-Skill --skill supervised-preference-training -a claude-code`. Or copy the skill folder (skills/repositories/repo-skills/openrlhf/sub-skills/supervised-preference-training in VectorSpaceLab/AREX-Skill) into .claude/skills/supervised-preference-training in your project. Claude Code loads it when a task matches its description.

How do I install Supervised Preference Training in Codex?

Run `npx skills add VectorSpaceLab/AREX-Skill --skill supervised-preference-training -a codex`. Or copy the skill folder (skills/repositories/repo-skills/openrlhf/sub-skills/supervised-preference-training in VectorSpaceLab/AREX-Skill) into .agents/skills/supervised-preference-training in your project. Codex loads it when a task matches its description.

Can I use Supervised Preference Training in Cursor, Gemini CLI or GitHub Copilot?

Cursor, Gemini CLI, GitHub Copilot and OpenCode also load SKILL.md folders. With the skills CLI, run `npx skills add VectorSpaceLab/AREX-Skill --skill supervised-preference-training -a cursor` (or -a gemini-cli, github-copilot or opencode for the others). To copy it by hand, put the folder in .cursor/skills/supervised-preference-training, .gemini/skills/supervised-preference-training, .github/skills/supervised-preference-training and .opencode/skills/supervised-preference-training in your project.

What does Supervised Preference Training need to run?

Going by SKILL.md and its folder, Supervised Preference Training needs Python for the scripts in its folder and the command-line tools its instructions call (python). Our summary lists: Python 3; Docker.

Does Supervised Preference Training access the network?

SKILL.md contains no URLs. Any network use would come from the scripts or tools the agent runs. This is read from the text; nothing was executed.

Is Supervised Preference Training safe to install?

Our automated static check of SKILL.md found no risky patterns, such as piping downloads into a shell, reading credential files or hidden Unicode. It is not a guarantee. The check reads SKILL.md only: the scripts in the folder are not scanned, so read them before running anything.

What licence does Supervised Preference Training use?

Supervised Preference Training is published under the Apache-2.0 licence (declared in SKILL.md). It allows redistribution, so the full SKILL.md is shown on this page.

How many tokens does Supervised Preference Training use?

About 1k tokens (SKILL.md is roughly 4.1k characters). Agents keep only the skill's name and description in context until a task matches; then they load SKILL.md in full. Its references folder adds about 5.2k tokens, read only when the agent opens those files.

What are the alternatives to Supervised Preference Training?

Skills that share tags, products or a category with Supervised Preference Training: Grpo (benchflow-ai/skillsbench, 1.8k stars), Slime User (yzlnew/infra-skills, 149 stars), Hugging Face LLM Trainer (huggingface/skills, 11k stars) and Train Rl (OpenPipe/ART, 11k stars). The comparison table on this page puts their stars, adoption, token cost, safety result and licence side by side.

Who maintains Supervised Preference Training?

VectorSpaceLab (a GitHub organization) maintains it in VectorSpaceLab/AREX-Skill, which has 328 GitHub stars. The repository holds 159 skills in this directory. The repository was last updated on September 3, 2026.

Source: VectorSpaceLab/AREX-Skill on GitHub. Facts on this page come from the repository at the commit we read; the author's words are quoted as theirs.