Deepseek Tiny V0.1
6-layer DeepSeek-V3 with Multihead Latent Attention (MLA) trained for research on shared subspaces in Transformer attention mechanisms.
Model Description
- Model Type: Transformer Decoder (DeepSeek-V3 based)
- Architecture: 6-layer decoder with Mixture of Experts
- Parameters: 16.26M
- Hidden Size: 256
- Attention Heads: 8
- Head Dimension: 32
- Sequence Length: 1,024 tokens
- Query Latent Dimension: 96
- Key-Value Latent Dimension: 64
Performance
- SST-2 Accuracy: 87.96%
- WikiText-103 Perplexity: 28.89
Research Context
This model is part of the shared-subspaces research project investigating the impact of shared output latent spaces in Transformer attention mechanisms.
Usage
import torch
from transformers import DeepseekV3ForCausalLM, AutoTokenizer
# Load model and tokenizer
model = DeepseekV3ForCausalLM.from_pretrained("ChrisMcCormick/deepseek-tiny-v0.1")
tokenizer = AutoTokenizer.from_pretrained("ChrisMcCormick/deepseek-tiny-v0.1")
# Generate text
inputs = tokenizer("The future of AI is", return_tensors="pt")
outputs = model.generate(**inputs, max_length=50, temperature=0.7)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
Training Details
- Pre-training Dataset: WikiText-103
- Fine-tuning Dataset: SST-2 (GLUE)
- Optimizer: AdamW
- Learning Rate: 5e-4 (pre-training), 5e-5 (fine-tuning)
- Weight Decay: 0.01 (pre-training), 0.05 (fine-tuning)
- Precision: bfloat16
- Compilation: torch.compile with inductor backend
- Training Steps: 12,500 (pre-training), 1,500 (fine-tuning)
Limitations
- Small scale model (16M parameters) intended for research purposes
- Trained on limited data compared to production models
- May require custom loading code for output subspace variants
Citation
@misc{mccormick2025sharedsubspaces,
title={Shared Subspaces in Transformer Attention: Investigating Output Latent Spaces},
author={McCormick, Chris},
year={2025},
howpublished={\url{https://github.com/chrisjmccormick/shared-subspaces}}
}
License
Apache 2.0
- Downloads last month
- 14