shivamfet/slm-125m-base

A ~125M-parameter Llama-architecture language model pretrained from scratch on a legal- and finance-heavy corpus (U.S. case law, SEC filings, and a slice of FineWeb-Edu). Built as a compact domain SLM; not instruction-tuned.

Model details

Parameters	~125.8M (tied embeddings)
Architecture	Llama (12 layers, 768 hidden, 12 heads)
Vocab	16,384 (byte-level BPE, trained on-corpus)
Context length	1024
Precision	bf16 autocast (fp32 master weights)

Training

• Corpus: ~2.4B unique tokens (case-law / SEC / FineWeb-Edu), deduplicated (MinHash LSH) and decontaminated against CaseHOLD / LexGLUE eval sets.
• Schedule: 4 epochs (~8.15B tokens seen) on 8xH100, cosine LR (6e-4 -> 6e-5), token-based warmup, fused AdamW.
• Result: validation perplexity 16.0 -> 8.35, monotonic, no divergence.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("shivamfet/slm-125m-base")
model = AutoModelForCausalLM.from_pretrained("shivamfet/slm-125m-base")

ids = tok("The plaintiff shall bear the burden of proving",
          return_tensors="pt", return_token_type_ids=False)
out = model.generate(**ids, max_new_tokens=80, do_sample=True,
                     top_k=50, temperature=0.8, pad_token_id=tok.pad_token_id)
print(tok.decode(out[0], skip_special_tokens=True))

Limitations

Base model only — no instruction tuning or RLHF. It can produce plausible-sounding but incorrect legal/financial text and must not be relied on for legal or financial advice. Reflects biases in its public training sources.