IMU-1 Base

This repository contains the IMU-1 Base model, a sample-efficient 430M parameter language model introduced in the paper IMU-1: Sample-Efficient Pre-training of Small Language Models.

IMU-1 is trained on 72B tokens and approaches the benchmark performance of models trained on 56× more data.

Model Details

Architecture

IMU-1 uses a validated recipe combining recent advances:

• QK-norm attention with learnable scale
• Per-head gating (sigmoid-based)
• Value residual learning
• LayerNorm scaling (depth-dependent)
• GQA (grouped query attention)
• SwiGLU activation
• RoPE positional encoding

Training

• Optimizer: NorMuon with cautious weight decay, muP parametrization
• Schedule: Three-stage WSD (Warmup-Stable-Decay)
• Post-processing: Checkpoint EMA (β=0.8)

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "thepowerfuldeez/imu1_base",
    trust_remote_code=True
)
tokenizer = AutoTokenizer.from_pretrained("thepowerfuldeez/imu1_base")

text = "The quick brown fox"
inputs = tokenizer(text, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(outputs[0]))

Note: This model uses custom modeling code. You must pass trust_remote_code=True when loading.

Benchmark Results

Training Stages

Resources

• Training Code: sample_efficient_gpt
• Stage 1 Data: 1218_imu1_base_stable_corpus
• Stage 2 Data: 1226_imu1_base_decay_corpus

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}, 
}

License

Apache 2.0