Skip to content

Latest commit

 

History

History

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

README.md

Hyper-parameters for GSM8K Finetuning on Qwen2.5-1.5b-Instruct

gsm8k_grpo.yaml uses 6e-6, retuned for the current FSDP FP32-master optimizer path.

Current FP32-Master Recipe

A seed-1 run on 8 NVIDIA A800 GPUs completed the official 10-epoch schedule with the following held-out evaluation results:

lr weight decay group size best eval reward final eval reward
6.0E-6 0.017 4 0.78412 0.77767

Historical Pre-FP32-Master Sweep

The results below were collected with BF16 parameter and optimizer-state storage. They are retained for reference and are not directly comparable with the current recipe.

lr weight decay group size max task_reward
1.70E-05 0.017 4 0.79570
1.30E-05 0.015 8 0.79355
1.50E-05 0.01 4 0.79043
1.50E-05 0.02 4 0.78984
1.00E-05 0.02 4 0.78311
1.00E-05 0.01 8 0.78066

Training Details

  • Devices: 8 Nvidia H800 GPUs
  • Optimizer: Adam
  • LR Scheduler: Constant
  • Gradient Clipping: 1.0
  • Max_new_tokens: 1024
  • Max_head_offpolicyness: 2
  • Training Time: ~35 minutes (batchsize 4), ~65 minutes (batchsize 8)