SLM-125M-Siva
A 125M parameter small language model trained on legal and financial text.
Architecture
- Type: LLaMA-style decoder-only transformer
- Parameters: ~125M (tied embeddings)
- Layers: 12 | Hidden: 768 | Heads: 12
- Context: 1024 tokens
- Vocab: 16,384 (BPE)
- Activation: SwiGLU with RoPE positional embeddings
Training Data
Legal-first corpus mix (~2B tokens):
- Case Law (~40%): US judicial opinions
- SEC Filings (~40%): Corporate financial disclosures
- FineWeb-Edu (~20%): High-quality web text
Usage
Ctrl click to launch VS Code Native REPL
Deployment
This model can be deployed as an API endpoint using Modal:
License
Apache 2.0