RohanAwhad

SciKnowEval RL — Beginner Run Guide (default prime-rl, GRPO only)

A step-by-step guide to reproduce cold-start GRPO training on SciKnowEval with prime-rl. By the end you will have RL-trained Qwen3-8B from scratch (no SFT warmstart) on multi-domain scientific MCQ and evaluated the result.

This guide is written from an actual run (Aug 2026) on an 8×H100 node. Everything here was verified to work. Companion to the

RohanAwhad

Reverse-Text RL — Beginner Run Guide

A step-by-step guide to reproduce the reverse-text RL experiment with prime-rl. By the end you will have fine-tuned Qwen3-0.6B with RL to reverse text, starting from a warm-started (SFT) checkpoint, and evaluated the result.

This guide is written from an actual run (Aug 2026) on an 8×H100 node. Everything here was verified to work.

---

0. What are we actually doing?

RohanAwhad

Tensor Parallelism Knowledge Compilation

Core Understanding

Tensor Parallelism (TP) shards individual tensors across devices, following the Megatron-LM pattern:

• Column-wise: Input projections (q/k/v, gate/up)
• Row-wise: Output projections (o_proj, down_proj)
• Sequence Parallel: Shards activations on sequence dimension for memory savings
• Loss Parallel: Keep logits sharded on vocab dimension for efficient cross-entropy

RohanAwhad

# /// script
# dependencies = [
#   "numpy",
# ]
# ///

# multi head latent attention

import numpy as np

RohanAwhad

#!/Users/rohan/miniconda3/bin/python

import asyncio
import dataclasses
import json
import openai
import os
import requests
import sys
import time

RohanAwhad

import json
import openai
from typing import Any
from jupyter_client.manager import KernelManager

kernel_manager: KernelManager = None
def execute_code_in_notebook(code: str) -> list[Any]:
  if not code: return []
  global kernel_manager
  if kernel_manager is None:

RohanAwhad

import base64
import cv2
import os
import openai
import whisper
import pyaudio
import wave
import uuid
import threading
from queue import Queue

RohanAwhad

from transformers import AutoTokenizer, AutoModel
import torch


# MODEL CKPT is downloaded from: "jinaai/jina-embeddings-v2-base-en"  # has context len of 8192
MODEL_CKPT = "/Users/rohan/3_Resources/ai_models/jina-embeddings-v2-base-en"

def recursive_splitter(text: str, separators: list[str], chunk_size: int) -> list[str]:
  if len(separators) == 0:
    words = text.strip().split(' ')

RohanAwhad

import dataclasses
import openai
import os
import re
import yaml

from pydantic import BaseModel, ValidationError
from typing import Optional

RohanAwhad

--[[

=====================================================================
==================== READ THIS BEFORE CONTINUING ====================
=====================================================================
========                                    .-----.          ========
========         .----------------------.   | === |          ========
========         |.-""""""""""""""""""-.|   |-----|          ========
========         ||                    ||   | === |          ========
========         ||   KICKSTART.NVIM   ||   |-----|          ========