rl-post-training

rl-post-training is a skill for Claude Code, Codex from benchflow-ai/skillsbench. It costs 126 tokens per session (1,411 once invoked), scanned A, original, Apache-2.0.

A troubleshooting guide for reinforcement-learning training of language models. It follows the path from generated answers and reward scores to model updates, explaining where learning can fail.

In plain words
What is it for?
Use it to diagnose GRPO, PPO, REINFORCE, or DPO post-training pipelines.
Why use it?
It helps locate problems such as constant rewards, ineffective advantages, stalled learning, NaN gradients, or incorrect text scoring.

Skill for Claude CodeCodex

Written for no agent in particular: nothing here depends on one.

Good fit Use it to diagnose GRPO, PPO, REINFORCE, or DPO post-training pipelines.

Compare 6 skills from other repositories ↓
Install with agentmods
npx agentmods add skills/benchflow-ai/skillsbench/rl-post-training
About the project

SkillsBench is a benchmark for measuring how effectively AI agents use modular skills—folders containing instructions, scripts, and resources—to complete specialized tasks. It helps researchers and developers evaluate both skill quality and agent behavior, including tasks that require combining multiple skills. The catalogue’s skills and instructions are evaluated as part of this workflow.

benchflow-ai/skillsbench · 1,764 stars · on GitHub · skillsbench.ai

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

Any agent
npx skills add benchflow-ai/skillsbench --skill rl-post-training
Clone the repo
git clone --depth 1 https://github.com/benchflow-ai/skillsbench

Made for: Claude Code, Codex.

Wrote this? Show the measurements

A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.

agentmods badge for rl-post-training

README.md
[![agentmods](https://agentmods.dev/badge/skills/benchflow-ai/skillsbench/rl-post-training/github.svg)](https://agentmods.dev/skills/benchflow-ai/skillsbench/rl-post-training)
Your own site
<a href="https://agentmods.dev/skills/benchflow-ai/skillsbench/rl-post-training"><img src="https://agentmods.dev/badge/skills/benchflow-ai/skillsbench/rl-post-training/github.svg" alt="Measured on agentmods" height="20"></a>

Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.

agentmods 80×15 button for rl-post-training

Your own site · 80×15
<a href="https://agentmods.dev/skills/benchflow-ai/skillsbench/rl-post-training"><img src="https://agentmods.dev/badge/skills/benchflow-ai/skillsbench/rl-post-training.svg" alt="Reviewed on agentmods" width="80" height="20"></a>
Per session 126 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 1,411 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. A grade says what 26 rules found in the file — not that it is safe. Third-party audits
  • NVIDIA SkillSpector pass 7 Sept 2026
How audits are shown
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5.1 $0.00126 $0.01411
Opus 5 $0.00063 $0.00705
Sonnet 5 $0.00025 $0.00282
Haiku 4.5 $0.00013 $0.00141

Measured 8d ago against content hash 3ff7177b0200, method: parsed. Prices are Anthropic first-party input rates as of 2026-09-11, from the pricing page.

Security

Grade A, and why

rl-post-training scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 8d ago.

The scan reads SKILL.md. This mod also ships 1 executable file (scripts/verify_pipeline.py), listed below but not scanned — reading those needs a real analyzer, not pattern matching.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

tasks/debug-trl-grpo/environment/skills/rl-post-training/SKILL.md · 89 lines

How it starts

The opening of the file, as written. The whole thing — 89 lines — stays where its author put it; the contents beside it link to each section on GitHub.

RL Post-Training — Concepts & Diagnostic Guide

Core Concepts

RL post-training optimizes a language model's policy using reward signals. The standard pipeline:

prompt → generate completions → score with reward → compute advantages → policy gradient update

Each stage has distinct failure modes. When a model "shows no improvement," the bug could be anywhere in this pipeline.

Diagnostic Methodology

When RL training produces no improvement, work through these stages in order. Each stage depends on the previous one being correct.

Stage 1: Verify Reward Signal

  • Are rewards non-constant? If all rewards are identical, there is no learning signal.
  • Do rewards correlate with completion quality? Spot-check decoded completions against their scores.
  • Is the reward function being called on the correct text? Check that decoding/stripping preserves the content the reward function needs to evaluate.

Stage 2: Verify Advantage Computation

  • Are advantages non-zero when rewards vary? If they collapse to ~0, the policy gradient vanishes.
  • Check the magnitude and dtype of every numerical-stability constant in the advantage path (additive epsilons, clipping bounds). Compare each to what the math requires.
  • Check the group size G. G ≤ 2 makes std either undefined or extremely noisy.

Stage 3: Verify Log-Probability Computation

  • Verify bounds: log-probs of valid tokens must be non-positive.
  • Compare your implementation against F.log_softmax on a small deterministic input — a numerical match rules out sign errors, wrong gathering axis, and off-by-one subtraction.
  • On a near-one-hot input, confirm the dominant token's log-prob is close to 0 (not close to the min).

Stage 4: Verify Loss Computation

  • Is the loss changing across steps? Flat loss suggests zero gradients upstream.
  • Log the KL term and the policy-gradient term separately. Either dominating the other is diagnostic.
  • Check the fraction of clipped samples. Near-100% clipping means the clip range is starving the signal.

Read the full file on GitHub · 89 lines

Files

What ships with it

3 files beside SKILL.md in the same directory: the scripts, references and assets a skill reads on demand. Not counted in the per-session cost; read them before you install if any of them is executable.

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 8d ago First seen · 89 lines · 126 tokens per session scan A 3ff7177b0200

Subscribe to this mod's changes

rl-post-training is a skill published in the GitHub repository benchflow-ai/skillsbench (1,764 stars, last pushed 1mo ago), licensed Apache-2.0. It adds 126 tokens to every session and 1,411 once invoked, about $0.0006 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-09-03.