A 125M-parameter Llama-style base language model , pretrained from scratch on a legal/financial corpus (US case law + SEC filings + a web slice, ~2.19B tokens). Pretraining: 10 epochs, 8xH100. Held-out validation perplexity: 8.36 . This is a base COMPLETER, not a chat model. Prompt it with the start of a sentence and it continues in the legal register. It speaks the register fluently but does...
--- license: other language: en tags: [legal, from-scratch, llama, small-language-model] --- # slm-125m-base A **125M-parameter Llama-style base language model**, pretrained **from scratch** on a legal/financial corpus (US case law + SEC filings + a web slice, ~2.19B tokens). - **Pretraining:** 10 epochs, 8xH100. **Held-out validation perplexity: 8.36**. - **This is a base COMPLETER, not a chat model.** Prompt it with the start of a sentence and it continues in the legal register. It speaks the register fluently but does not know facts (knowledge is capped at ~2 bits/param at this size). ```python from transformers import AutoModelForCausalLM, AutoTokenizer tok = AutoTokenizer.from_pretrained("thesreedath/slm-125m-base") model = AutoModelForCausalLM.from_pretrained("thesreedath/slm-125m-base") ids = tok("The plaintiff alleges that the defendant", return_tensors="pt").input_ids print(tok.decode(model.generate(ids, max_new_tokens=60, do_sample=True, temperature=0.8, min_new_tokens=40)[0])) ``` Full pipeline + fine-tuning code: https://github.com/Vizuara-AI-Lab/slm-125m-from-scratch
A 125M-parameter Llama-style base language model , pretrained from scratch on a legal/financial corpus (US case law + SEC filings + a web slice, ~2.19B tokens). Pretraining: 10 epochs, 8xH100. Held-out validation perplexity: 8.36 . This is a base COMPLETER,…
