token-averaging-avg_50m_k2_wide

FAIRC/token-averaging-avg_50m_k2_wide

Checkpoint dump from the token averaging research project.

• run name: avg_50m_k2_wide
• results tree: results

Contents

Loss logs

• loss_log.csv

Checkpoints

• checkpoints/final.pt

Loading a checkpoint

import torch
from huggingface_hub import hf_hub_download

path = hf_hub_download('FAIRC/token-averaging-avg_50m_k2_wide', 'checkpoints/final.pt')
state = torch.load(path, map_location='cpu', weights_only=False)
model.load_state_dict(state['model'])  # your OLMAveraged / OLMTransformerBody
print(state['step'], state['tokens_seen'], state['cumulative_flops'])

These are not Hugging Face transformers weights. Rebuild the architecture from config.json → model_config (or from experiments/chinchilla/model_configs.py in the source repo) and load the raw state_dict.

Architecture

{
  "d_model": 864,
  "n_heads": 8,
  "n_layers": 8,
  "context_len": 1024,
  "averaging_k": 2,
  "tie_embeddings": true,
  "lr": 0.0002,
  "warmup_steps": 2000,
  "target_tokens": 2000000000,
  "n_params_approx": 115085664
}