helios-314b-alpha

This repository contains JAX example code for loading and running the Helios-314B-Alpha open-weights model.

The Helios-314B-Alpha model is a trained version of the Grok-V1 open source model released by X.AI Corp.
We have fine-tuned the model to perform on crypto-related queries.
It achieves the following results on the evaluation set:

Loss: 0.0052
F1: 0.9969

Make sure to download the checkpoint and place the ckpt-0 directory in checkpoints

Then, run

pip install -r requirements.txt
python run.py

to test the code.

The script loads the checkpoint and samples from the model on a test input.

Due to the large size of the model (314B parameters), a machine with enough GPU memory is required to test the model with the example code. The implementation of the MoE layer in this repository is not efficient. The implementation was chosen to avoid the need for custom kernels to validate the correctness of the model.

Model Specifications

Helios is currently designed with the following specifications:

Parameters: 314B
Architecture: Mixture of 8 Experts (MoE)
Experts Utilization: 2 experts used per token
Layers: 64
Attention Heads: 48 for queries, 8 for keys/values
Embedding Size: 6,144
Tokenization: SentencePiece tokenizer with 131,072 tokens
Additional Features:
- Rotary embeddings (RoPE)
- Supports activation sharding and 8-bit quantization
Maximum Sequence Length (context): 8,192 tokens

License

The code and weights for the Helios-314B-Alpha model are licensed under the apache-2.0 open source license