legal-slm-125m-base
A 125M-parameter Llama-architecture small language model pretrained from scratch on a legal/financial corpus. Base (not instruction-tuned) model.
Results
- Validation perplexity: 10.67 (val_loss 2.3674), 1 epoch over 1.97B tokens.
Model
- 125,847,552 params — 12 layers / 768 hidden / 12 heads, context 1024, vocab 16,384 (custom byte-level BPE), tied embeddings, SwiGLU, RoPE.
Training data (~76% legal)
HFforLegal/case-law(US court opinions) — ~36%PleIAs/SEC(SEC filings) — ~40%HuggingFaceFW/fineweb-edu(educational web) — ~24%
Pipeline: deterministic cleaning (OCR gate, boilerplate/repetition/language filters) → MinHash near-dup + exact-dup removal → 13-gram decontamination against CaseHOLD/LexGLUE (so eval is uncontaminated) → 16K byte-level BPE → packed 1024-token windows.
Training
- 1 epoch, 3,753 steps, global batch 524,288 tokens, 2x H100 (DDP, bf16, torch.compile). AdamW (0.9/0.95, wd 0.1 on 2D params), grad-clip 1.0.
- LR: warmup to 6e-4 over 200M tokens, cosine to 6e-5 annealed over 1 epoch.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("VigneshwarKandhaiya/legal-slm-125m-base")
model = AutoModelForCausalLM.from_pretrained("VigneshwarKandhaiya/legal-slm-125m-base")
ids = tok("The plaintiff shall bear the burden of", return_tensors="pt").input_ids
print(tok.decode(model.generate(ids, max_new_tokens=40)[0], skip_special_tokens=True))
Limitations
Base model, English only, 125M params — expect fluent domain-styled text, not reliable facts or instruction-following. Not legal or financial advice.