unicosys-hypergraph / README.md
drzo's picture
chore: update unicosys hypergraph model from CI pipeline
425f10e verified
---
license: mit
tags:
- knowledge-graph
- hypergraph
- legal-evidence
- graph-neural-network
- unicosys
language:
- en
library_name: transformers
pipeline_tag: graph-ml
---
# Unicosys Hypergraph Knowledge Model
A trainable knowledge graph embedding model encoding the unified evidence
hypergraph for Case 2025-137857.
## Model Description
This model encodes a **unified hypergraph** linking financial transactions,
email communications, legal evidence, and entity relationships into a
single trainable knowledge representation.
### Architecture
| Component | Details |
|---|---|
| Node Embedding | 128-dim structural + 256-dim text |
| Hidden Dimension | 256 |
| Text Encoder | 2-layer Transformer, 4 heads |
| Graph Attention | 2-layer GAT, 4 heads |
| Link Predictor | 2-layer MLP with margin ranking loss |
| Total Parameters | **36,023,681** |
### Knowledge Graph Statistics
| Metric | Count |
|---|---|
| Total Nodes | 272,683 |
| Total Edges | 14,816 |
| Cross-Links | 3,976 |
| Entities | 16 |
| Emails | 199,553 |
| Financial Documents | 17,036 |
| Timeline Events | 54,346 |
| LEX Schemes | 13 |
| Legal Filings | 7 |
### Subsystems
| Subsystem | Nodes |
|---|---|
| Core (Entities) | 16 |
| Fincosys (Financial) | 72,935 |
| Comcosys (Communications) | 199,553 |
| RevStream1 (Evidence) | 146 |
| Ad-Res-J7 (Legal) | 33 |
## Training
The model can be fine-tuned on link prediction tasks:
```python
from model.unicosys_model import UnicosysHypergraphModel, UnicosysConfig
model = UnicosysHypergraphModel.from_pretrained("hyperholmes/unicosys-hypergraph")
# ... prepare training data ...
# model.forward(node_ids, node_type_ids, subsystem_ids, edge_index, edge_type_ids,
# pos_edge_index=pos, neg_edge_index=neg, labels=labels)
```
## Files
- `model.safetensors` β€” Model weights
- `config.json` β€” Model configuration
- `graph_data.safetensors` β€” Encoded graph tensors (nodes, edges)
- `tokenizer.json` β€” Character-level tokenizer for node labels
- `node_id_mapping.json` β€” Node ID string to integer index mapping
- `model_summary.json` β€” Compact statistics summary
## Source
Generated by the [Unicosys](https://github.com/hyperholmes/unicosys) intelligence pipeline.