StepLaw-N_268M-D_7.0B-LR4.883e-04-BS65536

Wandb Model Name: step2v2_0618_h1024_ffnh9552_numh16_numl8_lr4.883e-04_bs32_ti122070_mlr1e-5

This model is part of the StepLaw-N_268M-D_7.0B collection.

Model Specifications

Architecture

• Hidden size (H): 1024
• Feed-forward network size (FFN): 9552
• Attention heads: 16
• Layers: 8
• Parameter count: 268M

Training Parameters

• Learning rate (lr): 4.883e-04
• Batch size (bs): 65536
• Training iterations: 122070
• Training tokens (D): 8.0B

Model Description

StepLaw models are trained with various hyperparameter settings to enable research on scaling laws and hyperparameter optimization. This specific model was trained with learning rate 4.883e-04 and batch size 65536 for 122070 iterations, using a total of 8.0B training tokens.

Usage Example

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "StepLaw/StepLaw-N_268M-D_7.0B-LR4.883e-04-BS65536"
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True, use_fast=False)
model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True)

# Generate text
inputs = tokenizer("A long time ago in a galaxy far, far away", return_tensors="pt")
outputs = model.generate(**inputs, max_length=100)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))