---
license: apache-2.0
---
# 🧬 Evo-2 40B (BioNeMo) — Fine-tuned for Hepatocellular Carcinoma (HCC)
**Model Name:** `evo2-7b-bionemo-hcc-mouse`  
**Base Model:** [Evo-2 40B (Arc Institute × NVIDIA)](https://nvidia.github.io/bionemo-framework/models/evo2)  
**Framework:** NVIDIA BioNeMo / NeMo2 (Megatron-LM backend)  
**Fine-tuning Domain:** Hepatocellular Carcinoma (HCC)  
**Organism:** *Mus musculus* (GRCm39)  
**Chromosomes Used:** 1, 4, 6, 7, 8, 9, 11, 12, 17  
**Context:** Developed by the Brilliant team for the [EuroCC AI Hackathon 2025](https://www.openhackathons.org/s/siteevent/a0CUP000013Tp8f2AC/se000375)

## 🧠 Overview
This repository contains a **fine-tuned checkpoint** of the Evo-2 40B biomolecular foundation model, adapted to explore sequence patterns linked to **Hepatocellular Carcinoma (HCC)** using curated *Mus musculus* genomic data.  

Training leveraged the **NVIDIA BioNeMo** framework on multi-GPU H100 infrastructure during the **EuroCC AI Hackathon 2025**, focusing on **genome-scale representation learning** and **mutation-impact analysis** in cancer-related regions.

## ⚙️ Fine-tuning Dataset
| Property | Description |
|:--|:--|
| Source organism | *Mus musculus* |
| Genome build | GRCm39 |
| Chromosomes | 1, 4, 6, 7, 8, 9, 11, 12, 17 |
| Data type | DNA sequences (FASTA) |
| Task | Representation learning and sequence modeling for HCC pathways |
| Preprocessing | Standard BioNeMo Evo-2 pipeline (tokenization, masking, packing) |

## 🚀 Training Setup
| Parameter | Value |
|:--|:--|
| Base model | Evo-2 40B base |
| Framework | NVIDIA BioNeMo / NeMo2 |
| Precision | BF16 |
| GPUs | 16 × NVIDIA H100 |
| Tensor Parallel (TP) | 16 |
| Pipeline Parallel (PP) | 1 |
| Context Parallel (CP) | 1 |
| Micro-batch size | Tuned for memory efficiency |
| Dataset split | 90 % train / 5 % val / 5 % test |

## 🧩 Usage
```bash
python train_evo2.py   model.restore_from_path=./checkpoint   model.tensor_parallel_size=4   model.pipeline_model_parallel_size=2   trainer.precision=bf16
```

## 🧬 Background
### Evo-2
A genomic and protein language model jointly developed by **Arc Institute** and **NVIDIA**, trained on nearly 9 trillion biological tokens across many species and capable of context lengths up to 1 million tokens.  
- [NVIDIA Developer Blog – Understanding the Language of Life’s Biomolecules](https://developer.nvidia.com/blog/understanding-the-language-of-lifes-biomolecules-across-evolution-at-a-new-scale-with-evo-2)  
- [BioNeMo Framework Docs](https://nvidia.github.io/bionemo-framework/models/evo2)

### BioNeMo
NVIDIA’s AI platform for biomolecular modeling, built on NeMo 2 + Megatron-LM for scalable multi-GPU training and fine-tuning.

### EuroCC AI Hackathon 2025
Collaborative event (14–23 October 2025) promoting AI/HPC innovation, organized by EuroCC National Competence Centres in partnership with NVIDIA and OpenACC.  
- [Event Page](https://www.openhackathons.org/s/siteevent/a0CUP000013Tp8f2AC/se000375)

## 📜 License & Citation
This repository and checkpoint are released for **research and educational use** under the same terms as the original Evo-2 and BioNeMo frameworks (see LICENSE).

## ⚠️ Notes & Limitations
This fine-tuned version is **experimental**, trained on selected *Mus musculus* chromosomes related to HCC pathways.  
It is **not** intended for clinical or diagnostic use. Users must apply appropriate bioethical and biosafety standards in downstream applications.

## 💡 Acknowledgements
Special thanks to the EuroCC AI Hackathon mentors and NVIDIA BioNeMo team for guidance and compute resources, and to the Arc Institute for releasing the original Evo-2 foundation models.