lewtun

# pip install emoji
import argparse
from datasets import load_dataset
import emoji

def remove_emoji(text: str) -> str:
    return emoji.replace_emoji(text, replace='').strip()

def format_messages(x):
    emojis_found = False

lewtun

{# ───── defaults ───── #}
{%- if enable_thinking is not defined -%}
{%- set enable_thinking = true -%}
{%- endif -%}

{# ───── reasoning mode ───── #}
{%- if enable_thinking -%}
  {%- set reasoning_mode = "/think" -%}
{%- else -%}
  {%- set reasoning_mode = "/no_think" -%}

lewtun

from transformers import AutoModelForCausalLM, AutoTokenizer, GenerationConfig, set_seed
import time
import torch

set_seed(0)
device = "cuda"
model = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-1.5B",
    attn_implementation="flash_attention_2",
    torch_dtype="bfloat16"

lewtun

from datasets import load_dataset
from trl import GRPOConfig, GRPOTrainer
import random
"""Usage (on 8 x H100s):

pip install vllm==0.7.0 --extra-index-url https://download.pytorch.org/whl/cu121
pip install -e '.[dev]'

# DDP
accelerate launch --config_file examples/accelerate_configs/multi_gpu.yaml --num_processes 7 scratch/grpo_demo.py

lewtun

from typing import Optional

def extract_boxed_solution(text: str) -> Optional[str]:
    """
    Extracts the content of the last `\boxed{}` in a given LaTeX-style text.

    Args:
        text (str): The input string containing LaTeX-style content.

    Returns:

lewtun

# Copyright 2023 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,

lewtun

# flake8: noqa
# Copyright 2023 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software

lewtun

{
  "config_general": {
    "lighteval_sha": "?",
    "num_fewshot_seeds": 1,
    "override_batch_size": 4,
    "max_samples": null,
    "job_id": "",
    "start_time": 1163608.425196265,
    "end_time": 1173616.769654949,
    "total_evaluation_time_secondes": "10008.34445868386",

lewtun

"""
First install: pip install datasets pandas rich transformers

Usage:

# Loglikelihood evals
python view_details.py --filepath path/to/parquet/details 

# Generative evals
python view_details.py --filepath path/to/parquet/details --is_generative

lewtun

# This is a modified version of TRL's `SFTTrainer` example (https://github.com/huggingface/trl/blob/main/examples/scripts/sft_trainer.py), 
# adapted to run with DeepSpeed ZeRO-3 and Mistral-7B-V1.0. The settings below were run on 1 node of 8 x A100 (80GB) GPUs.
#
# Usage:
#   - Install the latest transformers & accelerate versions: `pip install -U transformers accelerate`
#   - Install deepspeed: `pip install deepspeed==0.9.5`
#   - Install TRL from main: pip install git+https://github.com/huggingface/trl.git
#   - Clone the repo: git clone github.com/huggingface/trl.git
#   - Copy this Gist into trl/examples/scripts
#   - Run from root of trl repo with: accelerate launch --config_file=examples/accelerate_configs/deepspeed_zero3.yaml --gradient_accumulation_steps 8 examples/scripts/sft_trainer.py