metadata

library_name: transformers
tags:
  - reasoning
license: apache-2.0
datasets:
  - d0rj/gsm8k-ru
language:
  - ru
base_model:
  - attn-signs/GPTR-8b-base

GPT Reasoner (V1)

[EN]
Reasoning model adapted for russian text generation.
Based on YandexGPT-pretrain -> GPTR-8b-base
[RU]
Модель рассуждений, адаптированная для генерации русскоязычного текста.
Построена на YandexGPT-pretrain -> GPTR-8b-base

Model Details / Детализация модели

[EN]
Reinforced GRPO version to invoke general reasoning capabilities.
This model can generate conditional and coherent chain-of-thought
[RU]
Версия RL GRPO для возможностей размышления и глубокого понимания запроса. Модель может генерировать когерентный текст русского языка на этой итерации.

Important:

[EN]
This is the first stage of reinforcement learning, don't expect the model to solve every mathematical problem. The training is ongoing. Still, this model is stable and can solve something now.
[RU]
Это первая стадия RL обучения, поэтому не стоит возлагать надежды на решения любой математической проблемы данной моделью. Обучение продолжается, данная версия модели скорее proof-of-concept, чем готовый математический ассистент. Несмотря на это, модель стабильна.

Further development

GRPO on Gromov dataset series

Model Description / Описание модели

Developed by: [Reisen Raumberg (Attention Signs team)]
Language(s) (NLP): [RU/EN]
SFT from model: [YandexGPT-5-lite-8B-pretrain]

Utilized HF.Accelerator
GPU hours: ~24h of NVIDIA A100

Для обучения использовался HuggingFace Accelerator
GPU часы: ~24h часа NVIDIA A100

Training Framework

GPTR was trained using MyLLM framework (by Attention Signs):
--==MyLLM==--

Model configuration (MyLLM Framework)

[model]
model_name_or_path = "attn-signs/GPTR-8-base"

[datasets]
dataset = "d0rj/gsm8k-ru"
problem_field = "question"
solution_field = "answer"
dataloader_num_workers = 2
test_size = 0.1
extract_hash = true

[run]
run_name = "rl-gptr-8"
report_to = "wandb"
logging_first_step = true
logging_steps = 1
save_strategy = "steps"
save_steps = 500
save_total_limit = 5
output_dir = "models/attn-signs-gptr-8-grpo"
project_name = "rl-gptr"

[training]
num_train_epochs = 1
per_device_train_batch_size = 2
learning_rate = 0.00001
bf16 = true
seed = 42
use_peft = true

[grpo]
use_vllm = true
num_generations = 2
max_completion_length = 2048
num_iterations = 1          # https://github.com/huggingface/trl/releases/tag/v0.16.0
scale_rewards = false       # should be default var
beta = 0.04                 # reference model beta in vllm
epsilon_high = 0.28         # Increasing upper bound epsilon leads to higher entropy during generation, promoting better exploration
preload_rm = false

[lora]
lora_target_modules = [
    "k_proj",
    "v_proj",
    "q_proj",
    "o_proj",
    "gate_proj",
    "up_proj",
    "down_proj",
]
lora_r = 32
lora_alpha = 64

[fusion]
use_liger = false
attn_implementation = "flash_attention_2"

[tokenizer]
eos_token =  "</s>"
pad_token = "<unk>"
chat_template = "{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% for message in messages %}{{'<s>' + message['role'] + '\n' + message['content'] + '</s>' + '\n'}}{% endfor %}{% if add_generation_prompt %}{{ '<s>assistant\n' }}{% endif %}"
force_chat_template = true
added_special_tokens = [
    "<think>",
    "</think>"
]
system_prompt = """
[MODE: Reflection]
"""

Rewards:

Equation structure reward
Correctness reward
Multilingual coherence reward
Strict chinese penalty
Format reward
Russian purity reward

Using the model / Как запустить?

repo = 'attn-signs/GPTR-8-v1'

model = AutoModelForCausalLM.from_pretrained(repo)
tokenizer = AutoTokenizer.from_pretrained(repo)

device = 'cuda' if torch.cuda.is_available() else 'cpu'
model.to(device)

user_prompt = '''
У уравнений x**2 + 2019ax + b = 0 и x**2 + 2019bx + a = 0 есть один общий корень. Чему может быть равен этот корень, если известно, что a != b?
'''
system_prompt = "[MODE: Reflection]"
messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": user_prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=4096
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]

print(response)