TernaryLM / README.md
nielsr's picture
nielsr HF Staff
Improve model card: add pipeline tag, move arxiv id, and link to code
07484e2 verified
|
raw
history blame
2.21 kB
metadata
datasets:
  - roneneldan/TinyStories
language:
  - en
license: apache-2.0
pipeline_tag: text-generation
tags:
  - efficient-llm
  - quantization
  - ternary
  - bitnet
  - pytorch
  - tinystories
  - language-modeling

TernaryLM-132M

TernaryLM is a 132M-parameter Transformer trained natively using ternary weights {-1, 0, +1} (approximately 1.58-bit effective precision).

Unlike post-training quantization (PTQ) methods that quantize pre-trained full-precision models, TernaryLM learns quantization-aware representations from scratch using straight-through estimators and adaptive per-layer scaling factors.

Resources

Architecture

  • Parameters: 132M
  • Layers: 12
  • Hidden Size: 768
  • Attention Heads: 12
  • Context Length: 512
  • Quantization: Native Ternary Training

Training

  • Dataset: TinyStories (~60k stories)
  • Optimizer: AdamW (betas=(0.9, 0.98))
  • Learning Rate: 3e-4
  • Scheduler: OneCycleLR
  • Epochs: 15
  • Hardware: Multi-GPU T4 setup (Kaggle)

Intended Use

Research on:

  • Efficient Transformers and architecture design.
  • Quantization-aware training (QAT) paradigms.
  • Deployment of LLMs in resource-constrained or edge environments.

Limitations

  • The model is a research prototype and is not instruction-tuned.
  • Pre-training was conducted on a relatively small dataset scale (TinyStories).

Citation

@misc{nargund2026ternarylmmemoryefficientlanguagemodeling,
      title={TernaryLM: Memory-Efficient Language Modeling via Native 1-Bit Quantization with Adaptive Layer-wise Scaling}, 
      author={Nisharg Nargund and Priyesh Shukla},
      year={2026},
      eprint={2602.07374},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2602.07374}, 
}