mistral-1L-tiny

A tiny single-layer 35.1M parameter Mistral model, with a hidden size of 512, and an MLP intermediate size of 1024. This model is trained on the roneneldan/TinyStories dataset. It achieves the following results on the evaluation set:

• Loss: 1.6868
• Accuracy: 0.5792

Model description

This work is inspired by the 21M parameter one-layer GPT-Neo of the Tiny Stories paper. Results reproduced to acquire high-frequency checkpoints for further analysis.

Intended uses & limitations

Analysis of feature dynamics and emergence in real-world language models.

Training procedure

Trained for 90171 steps, corresponding to ~2 hours on a single H100.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0006
• train_batch_size: 64
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• num_epochs: 3.0

Training results

Quite consistent English text generation.

Framework versions

• Transformers 4.38.1
• Pytorch 2.2.0+cu121
• Datasets 2.17.1
• Tokenizers 0.15.2