thepowerfuldeez/1226_imu1_base_decay_corpus

IMU-1 Stage 2 Training Corpus (Decay Phase)

Pre-tokenized training data for Stage 2 (decay phase) of IMU-1, a sample-efficient 430M parameter language model.

Dataset Details

Property	Value
Tokens	~28B
Format	Memory-mapped NumPy (.npy)
Tokenizer	SmolLM2-360M
Vocab size	49,152

Data Sources

Stage 2 uses tighter quality filters compared to Stage 1:

• DCLM-edu (higher threshold filtering)
• FineWeb-edu
• FineMath
• Curated high-quality sources

Download

huggingface-cli download thepowerfuldeez/1226_imu1_base_decay_corpus --repo-type=dataset

Usage with sample_efficient_gpt

# Clone training framework
git clone https://github.com/thepowerfuldeez/sample_efficient_gpt
cd sample_efficient_gpt

# Install dependencies
export UV_TORCH_BACKEND=auto
uv pip install setuptools uv_build maturin
uv sync

# Train Stage 2 (requires Stage 1 checkpoint)
uv run torchrun --nproc_per_node 8 train.py \
    --config configs/imu1_base.yaml \
    --config-key decay

Training Configuration (Stage 2)

Parameter	Value
Schedule	WSD (decay phase)
Iterations	100,000 (200k total)
Batch size	312
Context length	896
Muon LR	1.15e-2 → 25% min
Decay start	100k steps

Related Resources

• Model: thepowerfuldeez/imu1_base
• Stage 1 Data: thepowerfuldeez/1218_imu1_base_stable_corpus
• Training Code: sample_efficient_gpt

Citation

@misc{grigorev2026imu1sampleefficientpretrainingsmall,
      title={IMU-1: Sample-Efficient Pre-training of Small Language Models}, 
      author={George Grigorev},
      year={2026},
      eprint={2602.02522},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2602.02522}, 
}