Prime Lab Trainer

A Claude Code skill that builds, validates, and submits RL training environments on Prime Intellect Lab — from any HuggingFace dataset to a live GRPO training run, fully automated.

You describe the task. Claude writes the environment, validates it, pushes it to the Hub, and submits a hosted GRPO training run.

---

What It Does

Given a prompt like:

"Build an environment for cais/mmlu abstract algebra, use Qwen/Qwen3-30B-Instruct-2507, 1000 steps, and submit the training job."

Claude autonomously:

1. Inspects the dataset schema and answer format
2. Writes the reward function and environment package
3. Unit tests the reward function against auto-generated cases
4. Runs a real evaluation via Prime Inference
5. Validates the reward distribution is learnable
6. Pushes the environment to the Prime Hub
7. Submits a hosted GRPO training run

No GPU setup. No infrastructure config. One prompt.

---

Prerequisites

Accounts

• Prime Intellect account
• Claude Code (npm install -g @anthropic-ai/claude-code)

Install

pip install prime verifiers datasets

Login

prime login    # opens browser → authenticates everything

prime login covers inference, the environment hub, and training submission.

HuggingFace auth is not required for public datasets. Only needed if your dataset is private on HF, or you want to publish trained model checkpoints to HF Hub after training.

---

Quickstart

git clone https://github.com/abideen/prime-lab-trainer
cd prime-lab-trainer
pip install prime verifiers datasets
prime login
claude

Then in Claude:

Build an RL training environment for the pubmed_qa dataset and submit training.

---

How It Works

This project is a Claude Code skill — a structured instruction set in SKILL.md that Claude reads automatically when you open the project. Claude follows a strict 7-step workflow, validating at every stage before proceeding.

Step 0  Preflight Check        Verify environment, credentials, tools
Step 1  Inspect Dataset        Learn schema, columns, answer format
Step 2  Write Environment      Reward function + 3-file package
Step 3  Validate Reward        Unit tests against auto-generated cases
Step 4  Local Eval             Real model completions via Prime Inference
Step 5  Distribution Check     Statistical validation of reward signal
Step 6  Push to Hub            Publish environment to Prime Hub
Step 7  Submit Training        GRPO training run on hosted GPUs

Claude never skips a step. If any step fails, it stops and fixes the issue before continuing.

---

The 7-Step Workflow

Step 0 — Preflight

Verifies all prerequisites before touching any code.

python scripts/preflight.py
# or auto-install missing packages:
python scripts/preflight.py --install

Checks: Python ≥ 3.10, datasets, verifiers, prime CLI, Prime login.

---

Step 1 — Inspect Dataset

Loads the dataset and reports its exact schema before writing any code.

python scripts/inspect_dataset.py pubmed_qa
python scripts/inspect_dataset.py openai/gsm8k main train

Output includes column names, types, sample rows, and an auto-detected answer format (GSM8K ####, numeric, boolean, multi-choice, or free text).

---

Step 2 — Write the Environment

Creates three files in environments/my_env/:

environments/my_env/
├── my_env.py        ← reward function + load_environment()
├── pyproject.toml   ← package metadata and dependencies
└── README.md        ← required for Hub display

Environments use the verifiers library. See references/environment_guide.md for complete patterns: classification, multi-choice, judge-scored free text, multi-turn, and more.

---

Step 3 — Unit Test the Reward

Auto-generates test cases from the environment's own dataset and runs sanity checks.

python scripts/test_reward.py environments/my_env/my_env.py
python scripts/test_reward.py environments/my_env/my_env.py --verbose

Sanity checks:

• Gold completions score higher than wrong completions
• At least one gold reward > 0.5
• At least one wrong reward < 0.5
• Not all rewards identical

✓ All sanity checks passed. Reward function looks correct.
  Proceed to: prime eval run <env-name> -m  -n 50

---

Step 4 — Local Evaluation

Runs real model completions against the environment using Prime Inference.

prime env install my-env --with pip
prime eval run my-env -m qwen/qwen3-vl-30b-a3b-instruct -n 50

Uses your prime login credentials — no additional API key needed. See prime inference models for the full list of available models.

---

Step 5 — Reward Distribution Check

Statistically validates that the reward signal is learnable before committing to a training run.

python scripts/check_reward_distribution.py

Check	Pass Condition	Failure Means
Sample count	n ≥ 20	Too few samples
Not all zero	mean > 0.05	Reward function broken
Not all ones	mean < 0.95	Trivially easy / reward hacked
Variance	std > 0.05	Training signal too weak
Format reward	> 0.3	Model can't follow format

All 5 must pass before pushing to hub.

---

Step 6 — Push to Hub

Publishes the environment to the Prime Environments Hub.

prime env push my-env

Verify at: https://app.primeintellect.ai/dashboard/environments

---

Step 7 — Submit Training

Generates a TOML config and submits a hosted GRPO training run.

python scripts/submit_training.py --env your-username/my-env
# or with flags:
python scripts/submit_training.py --env your-username/my-env \
  --model Qwen/Qwen3-30B-Instruct-2507 \
  --steps 1000 \
  --yes

GPU allocation is fully managed — no instance config needed.

Monitor your run:

prime rl list
prime rl logs <run-id> -f
prime rl metrics <run-id>

Dashboard: https://app.primeintellect.ai/dashboard/training

---

Fast Path — Existing Hub Environment

If the environment already exists on the Prime Hub, skip Steps 1–6.

First, disambiguate — both HF datasets and Prime environments use owner/name format:

prime env info owner/name
# exit 0  → Prime environment → use fast path
# exit 1  → HF dataset or not found → follow Steps 1–7

Then submit directly:

python scripts/submit_training.py --env zain/OpenMed_PubMedQA

---

Project Structure

prime-lab-trainer/
│
├── SKILL.md                        ← Claude Code skill definition (agent instructions)
│
├── scripts/
│   ├── preflight.py                ← Step 0: verify setup
│   ├── inspect_dataset.py          ← Step 1: inspect HF dataset schema
│   ├── test_reward.py              ← Step 3: unit test reward functions
│   ├── check_reward_distribution.py← Step 5: validate reward distribution
│   └── submit_training.py          ← Step 7: generate TOML + submit
│
├── references/
│   ├── environment_guide.md        ← Full environment authoring reference
│   └── training_guide.md           ← Full training config reference
│
├── environments/                   ← Your environments go here
│   └── my_env/
│       ├── my_env.py
│       ├── pyproject.toml
│       └── README.md
│
└── configs/
    └── rl/                         ← Auto-generated TOML configs
        └── my-env.toml

---

---

References

• Prime Intellect Docs
• Verifiers Library
• Prime Environments Hub
• Prime Training Dashboard
• Prime Inference Models