slm-125m-base

A 125.8M-parameter Llama-architecture language model, trained from scratch on a legal/financial corpus.

Data

Trained on 2.19B tokens: US case law (40%), SEC filings (40%), and FineWeb-Edu web text (20%). See the project's Agent_brief.md for the full data pipeline (cleaning, dedup, decontamination).

Architecture

• vocab_size: 16384
• hidden_size: 768
• num_hidden_layers: 12
• num_attention_heads: 12
• max_position_embeddings: 1024

Training

Pretrained on 8xH100 via Modal, budget-capped at $40.