mirror of
https://github.com/open-thought/reasoning-gym.git
synced 2026-04-22 16:49:06 +00:00
56ce2e79a7
* v0 * 2 gpu setup * improve parsing from yaml * update yaml dataset example * remove restriction on flash attn * more comments * first version of the readme * pin torch * simplify requirements * just flash attn * use set env instead * simpler set env * readme * add wandb project to setup * update template * update model id * post init to capture the config and weight * extract metadata * update config * update dataset config * move env for wandb project * pre-commit * remove qwen-math from training * more instructions * unused import * remove trl old * warmup ratio * warmup ratio * change model id * change model_id * add info about CUDA_VISIBLE_DEVICES
52 lines
1 KiB
YAML
52 lines
1 KiB
YAML
# Reasoning Gym configs
|
|
dataset_size: 20000
|
|
developer_prompt: DeepSeekZero
|
|
developer_role: system
|
|
datasets:
|
|
simple_equations:
|
|
weight: 1
|
|
complex_arithmetic:
|
|
weight: 1
|
|
config:
|
|
min_real: -20
|
|
max_real: 20
|
|
|
|
|
|
# Model configs from trl
|
|
model_name_or_path: Qwen/Qwen2.5-1.5B-Instruct
|
|
attn_implementation: flash_attention_2
|
|
|
|
# GRPO trainer configs from trl
|
|
bf16: true
|
|
use_vllm: true
|
|
vllm_device: cuda:1
|
|
vllm_gpu_memory_utilization: 0.9
|
|
log_level: info
|
|
gradient_accumulation_steps: 1
|
|
gradient_checkpointing: true
|
|
gradient_checkpointing_kwargs:
|
|
use_reentrant: false
|
|
logging_first_step: true
|
|
logging_steps: 1
|
|
logging_strategy: steps
|
|
learning_rate: 1e-06
|
|
lr_scheduler_type: constant_with_warmup
|
|
lr_scheduler_kwargs:
|
|
num_warmup_steps: 10
|
|
max_prompt_length: 512
|
|
max_completion_length: 2048
|
|
max_steps: 100
|
|
num_generations: 8
|
|
num_train_epochs: 1
|
|
overwrite_output_dir: true
|
|
per_device_train_batch_size: 8
|
|
report_to:
|
|
- wandb
|
|
|
|
save_strategy: steps
|
|
save_steps: 50
|
|
save_total_limit: 5
|
|
|
|
seed: 42
|
|
temperature: 0.6
|
|
warmup_ratio: 0.1
|