Supervised Preference Training
VectorSpaceLab/AREX-Skill
Build and review OpenRLHF supervised/preference training plans for SFT, reward models, DPO, IPO, and cDPO.
Reference for the GRPO (Group Relative Policy Optimization) algorithm.
Install Grpo
$ npx skills add benchflow-ai/skillsbench --skill grpo -a claude-codeProject install by default; add -g for ~/.claude/skills/.
$ gh skill install benchflow-ai/skillsbench grpo --agent claude-codeProject scope by default; add --scope user for a personal install. Needs GitHub CLI 2.90.0 or later (public preview).
$ git clone --depth 1 https://github.com/benchflow-ai/skillsbench.git skills-src && mkdir -p .claude/skills && cp -r skills-src/tasks/debug-trl-grpo/environment/skills/grpo .claude/skills/grpo && rm -rf skills-srcUse ~/.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/
Install the "grpo" agent skill from https://github.com/benchflow-ai/skillsbench/tree/main/tasks/debug-trl-grpo/environment/skills/grpo into .claude/skills/grpo/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "grpo", then confirm the skill loads.Claude Code copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
$skill-installer install https://github.com/benchflow-ai/skillsbench/tree/main/tasks/debug-trl-grpo/environment/skills/grpoType this inside Codex. $skill-installer <name> installs a curated skill from openai/skills. The installer writes to $CODEX_HOME/skills (default ~/.codex/skills). Restart Codex if the skill does not show up.
$ npx skills add benchflow-ai/skillsbench --skill grpo -a codexProject install goes to .agents/skills/; add -g for ~/.codex/skills/.
$ gh skill install benchflow-ai/skillsbench grpo --agent codexProject scope by default (.agents/skills/); add --scope user for a personal install.
$ git clone --depth 1 https://github.com/benchflow-ai/skillsbench.git skills-src && mkdir -p .agents/skills && cp -r skills-src/tasks/debug-trl-grpo/environment/skills/grpo .agents/skills/grpo && rm -rf skills-srcUse ~/.agents/skills/ instead of .agents/skills for a personal install.
Codex skills documentation · loads skills from .agents/skills/
Install the "grpo" agent skill from https://github.com/benchflow-ai/skillsbench/tree/main/tasks/debug-trl-grpo/environment/skills/grpo into .agents/skills/grpo/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "grpo", then confirm the skill loads.Codex copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
$ npx skills add benchflow-ai/skillsbench --skill grpo -a cursorProject install goes to .agents/skills/; add -g for ~/.cursor/skills/.
$ gh skill install benchflow-ai/skillsbench grpo --agent cursorProject scope by default (.agents/skills/); add --scope user for a personal install.
$ git clone --depth 1 https://github.com/benchflow-ai/skillsbench.git skills-src && mkdir -p .cursor/skills && cp -r skills-src/tasks/debug-trl-grpo/environment/skills/grpo .cursor/skills/grpo && rm -rf skills-srcUse ~/.cursor/skills/ instead of .cursor/skills for a personal install.
Cursor skills documentation · loads skills from .cursor/skills/, .agents/skills/, .claude/skills/, .codex/skills/
Install the "grpo" agent skill from https://github.com/benchflow-ai/skillsbench/tree/main/tasks/debug-trl-grpo/environment/skills/grpo into .cursor/skills/grpo/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "grpo", then confirm the skill loads.Cursor copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
$ gemini skills install https://github.com/benchflow-ai/skillsbench.git --path tasks/debug-trl-grpo/environment/skills/grpo--scope user (default) or --scope workspace; --path is the subfolder of the repo that holds the skill; --consent skips the security confirmation prompt.
$ npx skills add benchflow-ai/skillsbench --skill grpo -a gemini-cliProject install goes to .agents/skills/; add -g for ~/.gemini/skills/.
$ gh skill install benchflow-ai/skillsbench grpo --agent gemini-cliProject scope by default (.agents/skills/); add --scope user for a personal install.
$ git clone --depth 1 https://github.com/benchflow-ai/skillsbench.git skills-src && mkdir -p .gemini/skills && cp -r skills-src/tasks/debug-trl-grpo/environment/skills/grpo .gemini/skills/grpo && rm -rf skills-srcUse ~/.gemini/skills/ instead of .gemini/skills for a personal install, then run /skills reload.
Gemini CLI skills documentation · loads skills from .gemini/skills/, .agents/skills/
Install the "grpo" agent skill from https://github.com/benchflow-ai/skillsbench/tree/main/tasks/debug-trl-grpo/environment/skills/grpo into .gemini/skills/grpo/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "grpo", then confirm the skill loads.Gemini CLI copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
$ gh skill install benchflow-ai/skillsbench grpoInstalls for Copilot at project scope by default; add --scope user for a personal install. Preview a skill first with gh skill preview. Needs GitHub CLI 2.90.0 or later (public preview).
$ npx skills add benchflow-ai/skillsbench --skill grpo -a github-copilotProject install goes to .agents/skills/; add -g for ~/.copilot/skills/.
$ git clone --depth 1 https://github.com/benchflow-ai/skillsbench.git skills-src && mkdir -p .github/skills && cp -r skills-src/tasks/debug-trl-grpo/environment/skills/grpo .github/skills/grpo && rm -rf skills-srcUse ~/.copilot/skills/ instead of .github/skills for a personal install. Commit .github/skills so cloud agent and code review can use it.
GitHub Copilot skills documentation · loads skills from .github/skills/, .claude/skills/, .agents/skills/
Install the "grpo" agent skill from https://github.com/benchflow-ai/skillsbench/tree/main/tasks/debug-trl-grpo/environment/skills/grpo into .github/skills/grpo/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "grpo", then confirm the skill loads.GitHub Copilot copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
$ npx skills add benchflow-ai/skillsbench --skill grpo -a opencodeOpenCode documents no install command of its own. Project install goes to .agents/skills/; add -g for ~/.config/opencode/skills/.
$ gh skill install benchflow-ai/skillsbench grpo --agent opencodeProject scope by default (.agents/skills/); add --scope user for a personal install.
$ git clone --depth 1 https://github.com/benchflow-ai/skillsbench.git skills-src && mkdir -p .opencode/skills && cp -r skills-src/tasks/debug-trl-grpo/environment/skills/grpo .opencode/skills/grpo && rm -rf skills-srcUse ~/.config/opencode/skills/ instead of .opencode/skills for a personal install.
OpenCode skills documentation · loads skills from .opencode/skills/, .claude/skills/, .agents/skills/
Install the "grpo" agent skill from https://github.com/benchflow-ai/skillsbench/tree/main/tasks/debug-trl-grpo/environment/skills/grpo into .opencode/skills/grpo/ in this project. Copy the whole folder (SKILL.md and every file beside it), keep the folder name "grpo", then confirm the skill loads.OpenCode copies the folder itself, the same result as the manual copy. Check what it changed before you commit it.
Facts
- Skill name
grpo- GitHub stars
- 1.8k
- Token cost
- ~1.1k tokens
- SKILL.md length
- 383 words
- Files
- 3 (incl. references)
- Skills in repo
- 178
- Repo updated
- First seen
- Licence
- Apache-2.0
At a glance
Reference for the GRPO (Group Relative Policy Optimization) algorithm.
- Verifying a GRPO training pipeline — covers the mathematical formulation (group-relative advantages
- SKILL.md covers Overview, Training Loop, Advantage Estimation and Log-Probability Computation, plus 3 more sections
- Instructions only: no scripts, shell commands, URLs or credentials in SKILL.md
- Clipped surrogate loss
What it does
Grpo is an agent skill from benchflow-ai/skillsbench. Reference for the GRPO (Group Relative Policy Optimization) algorithm. Use when implementing, debugging, or verifying a GRPO training pipeline — covers the mathematical formulation (group-relative advantages, clipped surrogate loss, KL penalty), the training loop (generate → score → advantage → loss), log-probability computation, advantage estimation, and relationship to PPO/REINFORCE.
Its SKILL.md is about 1.1k tokens, which your agent loads only when the skill is triggered. The skill folder holds 3 other files, including reference files (for example `references/grpo-algorithm.md` and `references/grpo-trainer-internals.md`).
It sits in AI & LLM Engineering, covering Fine-tuning, Deep learning and Reinforcement learning. The repository describes itself as: SkillsBench evaluates how well skills work and how effective agents are at using them. The licence is Apache-2.0.
When your agent uses it
- Verifying a GRPO training pipeline — covers the mathematical formulation (group-relative advantages
- Clipped surrogate loss
- The training loop (generate → score → advantage → loss)
- Log-probability computation
Example prompts
- “/grpo”
What it can do on your machine
Read from SKILL.md and the folder at commit 9a1f4dd. 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
No scripts in the folder and no shell commands in SKILL.md.
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
Grpo loads about 1.1k tokens when it runs, and up to ~3.3k if it reads all its reference files. Until then it costs about 98 tokens; SKILL.md has 383 words of instructions outside code blocks.
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); files beside SKILL.md are not scanned.
SKILL.md
The full file from benchflow-ai/skillsbench at commit 9a1f4dd, republished under its Apache-2.0 licence (© benchflow-ai). 383 words, ~1,106 tokens.
Download SKILL.mdSave it as
.claude/skills/grpo/SKILL.md (or your agent's skills folder). This skill also uses 2 other files; get the full folder from GitHub.- name
- grpo
- description
- Reference for the GRPO (Group Relative Policy Optimization) algorithm. Use when implementing, debugging, or verifying a GRPO training pipeline — covers the mathematical formulation (group-relative advantages, clipped surrogate loss, KL penalty), the training loop (generate → score → advantage → loss), log-probability computation, advantage estimation, and relationship to PPO/REINFORCE.
GRPO Algorithm Reference
Overview
Group Relative Policy Optimization (GRPO) is a policy gradient method that eliminates the need for a learned critic by computing advantages from group statistics. For each prompt, multiple completions are sampled and their rewards are normalized within the group to produce advantages.
Key insight: Instead of training a value function to estimate V(s), GRPO uses the mean reward of the group as the baseline. This removes the critic entirely, reducing memory and avoiding value function approximation errors.
Training Loop
Each step follows this pipeline:
1. Sample G completions per prompt from current policy
2. Decode completions to text
3. Score completions with reward function
4. Compute group-relative advantages
5. Compute per-token log-probs (current policy + reference policy)
6. Compute clipped surrogate loss with KL penalty
7. Backpropagate and updateAdvantage Estimation
Rewards are normalized within each group of G completions for the same prompt:
mu = mean(r_1, ..., r_G)
sigma = std(r_1, ..., r_G)
A_i = (r_i - mu) / (sigma + epsilon)epsilon is a small constant (typically 1e-4 to 1e-8) for numerical stability only. It prevents division by zero when all rewards in a group are identical.
Properties:
- Advantages within each group sum to approximately zero;
- High-reward completions get positive advantages, low-reward get negative;
- The learning signal vanishes if epsilon is too large (dominates denominator) or if rewards are constant. For example, if all rewards in a group are identical (or nearly identical), then causing all advantages to collapse to 0;
Log-Probability Computation
Per-token log probabilities via log-softmax:
log_prob(token_i) = logit(token_i) - logsumexp(logits)Critical invariant: log_prob <= 0 always, since it's the log of a probability in (0, 1].
Sequence-level log probability: log_pi(y|x) = sum_t log_pi(t_j | x, t_<j)
Show full SKILL.md (170 more words)Show less
Loss Function
The GRPO loss combines a clipped surrogate objective with a KL divergence penalty:
ratio = exp(log_pi_theta(y|x) - log_pi_old(y|x))
L_clip = min(ratio * A, clip(ratio, 1-eps, 1+eps) * A)
L_kl = beta * KL(pi_theta || pi_ref)
loss = -E[L_clip] + L_kl| Component | Purpose |
|---|---|
ratio | How much the policy has changed from the generation policy |
| Clipping | Prevents destructively large policy updates |
beta * KL | Keeps the policy close to the reference (prevents degeneration) |
Key Hyperparameters
| Parameter | Typical Range | Effect |
|---|---|---|
num_generations (G) | 4–16 | More = lower variance advantages, higher compute cost |
beta | 0.01–0.1 | Higher = more conservative updates (closer to reference) |
epsilon (clip) | 0.1–0.2 | Narrower = more conservative updates |
epsilon (advantage) | 1e-8–1e-4 | Must be small; only for numerical stability |
learning_rate | 1e-7–5e-6 | Much lower than SFT; RL is sensitive to LR |
Available References
| File | Contents | When to load |
|---|---|---|
references/grpo-algorithm.md | Full mathematical formulation with notation table, step-by-step derivations, comparison to PPO and REINFORCE | When you need to verify whether a specific implementation detail matches the algorithm specification |
references/grpo-trainer-internals.md | GRPOTrainer implementation from popular frameworks (TRL): method-by-method breakdown, data flow diagram, and how each algorithm step maps to code | When tracing bugs through a GRPOTrainer implementation or understanding how the algorithm maps to specific code paths |
© benchflow-ai, 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 2 other files (references) in tasks/debug-trl-grpo/environment/skills/grpo of benchflow-ai/skillsbench.
- SKILL.md
- references/grpo-algorithm.md
- references/grpo-trainer-internals.md
Open the folder on GitHubat commit 9a1f4dd
Compare with similar skills
Grpo 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.
| Skill | Stars | Used in | Tokens | Auto-check | Licence | Repo updated |
|---|---|---|---|---|---|---|
| Grpo this skillbenchflow-ai/skillsbench | 1.8k | — | ~1.1k | Automated safety check: Pass | Apache-2.0 | |
| Supervised Preference TrainingVectorSpaceLab/AREX-Skill | 328 | — | ~1k | Automated safety check: Pass | Apache-2.0 | |
| Slime Useryzlnew/infra-skills | 149 | — | ~3.2k | Automated safety check: Pass | None | |
| Hugging Face LLM Trainerhuggingface/skills | 11k | 3 repos | ~7.2k | Automated safety check: Pass | Apache-2.0 | |
| Train RlOpenPipe/ART | 11k | — | ~2.4k | Automated safety check: Pass | Apache-2.0 | |
| Qwopus27b Rl TrainingR6410418/Jackrong-llm-finetuning-guide | 1.7k | — | ~830 | Automated safety check: Pass | Apache-2.0 |
Similar skills
Slime User
yzlnew/infra-skills
Guide for using SLIME (LLM post-training framework for RL Scaling).
- Official
Hugging Face LLM Trainer
huggingface/skills
Trains or fine-tunes language and vision models with TRL or Unsloth on Hugging Face Jobs cloud GPUs, then converts the results to GGUF.
Train Rl
OpenPipe/ART
RL training reference for the ART framework. An agent skill from OpenPipe/ART.
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.
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.
More from benchflow-ai/skillsbench
All 178 skills in this repoLean4 Memories
benchflow-ai/skillsbench
This skill should be used when working on Lean 4 formalization projects to maintain persistent memory of successful proof patterns, failed approaches, project conventions, and user preferences…
Senior Data Engineer
benchflow-ai/skillsbench
World-class data engineering skill for building scalable data pipelines, ETL/ELT systems, real-time streaming, and data infrastructure.
Ac Branch Pi Model
benchflow-ai/skillsbench
AC branch pi-model power flow equations (P/Q and |S|) with transformer tap ratio and phase shift, matching acopf-math-model.md and MATPOWER branch fields.
Civ6lib
benchflow-ai/skillsbench
Civilization 6 district mechanics library. An agent skill from benchflow-ai/skillsbench.
D3 Visualization
benchflow-ai/skillsbench
Build deterministic, verifiable data visualizations with D3.js (v6).
Dc Power Flow
benchflow-ai/skillsbench
DC power flow analysis for power systems. An agent skill from benchflow-ai/skillsbench.
Related
Categories
Questions about Grpo
What does Grpo do?
Reference for the GRPO (Group Relative Policy Optimization) algorithm. Grpo is an agent skill from benchflow-ai/skillsbench. Reference for the GRPO (Group Relative Policy Optimization) algorithm.
When should I use Grpo?
Grpo fits situations like: verifying a GRPO training pipeline — covers the mathematical formulation (group-relative advantages; clipped surrogate loss; the training loop (generate → score → advantage → loss); log-probability computation.
How do I install Grpo in Claude Code?
Run `npx skills add benchflow-ai/skillsbench --skill grpo -a claude-code`. Or copy the skill folder (tasks/debug-trl-grpo/environment/skills/grpo in benchflow-ai/skillsbench) into .claude/skills/grpo in your project. Claude Code loads it when a task matches its description.
How do I install Grpo in Codex?
Run `npx skills add benchflow-ai/skillsbench --skill grpo -a codex`. Or copy the skill folder (tasks/debug-trl-grpo/environment/skills/grpo in benchflow-ai/skillsbench) into .agents/skills/grpo in your project. Codex loads it when a task matches its description.
Can I use Grpo 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 benchflow-ai/skillsbench --skill grpo -a cursor` (or -a gemini-cli, github-copilot or opencode for the others). To copy it by hand, put the folder in .cursor/skills/grpo, .gemini/skills/grpo, .github/skills/grpo and .opencode/skills/grpo in your project.
What does Grpo need to run?
SKILL.md names no scripts, command-line tools or credentials: Grpo is instructions for the agent only.
Does Grpo 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 Grpo 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. Review the folder before installing.
What licence does Grpo use?
Grpo is published under the Apache-2.0 licence (the repository's licence). It allows redistribution, so the full SKILL.md is shown on this page.
How many tokens does Grpo use?
About 1.1k tokens (SKILL.md is roughly 4.4k 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 2.2k tokens, read only when the agent opens those files.
What are the alternatives to Grpo?
Skills that share tags, products or a category with Grpo: Supervised Preference Training (VectorSpaceLab/AREX-Skill, 328 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 Grpo?
benchflow-ai (a GitHub organization) maintains it in benchflow-ai/skillsbench, which has 1,832 GitHub stars. The repository holds 178 skills in this directory. The repository was last updated on July 23, 2026.
Source: benchflow-ai/skillsbench on GitHub. Facts on this page come from the repository at the commit we read; the author's words are quoted as theirs.