fblgit / miniclaus-qw1.5B-UNAMGS-GRPO

This version is RL with GRPO on GSM8k for 1400 steps using this code: # traingrpo.py import re import torch from datasets import loaddataset, Dataset from transformers import AutoTokenizer, AutoModelForCausalLM from peft import LoraConfig from trl import GRPOConfig, GRPOTrainer