slm125live-base

SLM125 — a 125M-parameter legal/financial base language model, trained from scratch

SLM125 is a small (~125.8M parameter), decoder-only transformer pretrained from scratch on a legal- and finance-heavy corpus. It is a base model — it completes text, it is not instruction-tuned or chat-tuned. Give it a sentence to continue ("The plaintiff argued that ..."), not a question.

🔗 Try it live: SLM125 Playground

Model Details

• Architecture: Llama-style decoder-only transformer (RoPE, RMSNorm, SwiGLU MLP, tied embeddings)
• Parameters: ~125.8M
• Hidden size: 768
• Layers: 12
• Attention heads: 12 (head dim 64, standard multi-head attention — no GQA)
• Intermediate (MLP) size: 3,072
• Context length: 1,024 tokens
• Vocabulary: 16,384 tokens (custom-trained tokenizer)
• RoPE theta: 10,000
• Precision: trained and served in fp32/bf16 on Modal

Training Data

Pretrained on a cleaned, deduplicated, decontaminated corpus of ~2.19B tokens, mixed "legal-first" from three public, streamed HuggingFace datasets:

The two legal sources were taken in full (they cap out around 2B tokens combined); a smaller web slice was added on top, landing at roughly a 40/40/20 split — about 78% legal/financial text overall, not the originally-planned 70/20/10.

How to Get Started

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "srinivasch87/slm125live-base"
tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo)

prompt = "The plaintiff argued that"
inputs = tokenizer(prompt, return_tensors="pt")
output = model.generate(
    **inputs,
    max_new_tokens=128,
    do_sample=True,
    temperature=0.8,
    top_p=0.95,
    top_k=50,
)
print(tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Intended Use & Limitations

• This is a base completion model, not an instruction-following assistant. It will not reliably answer questions, follow chat-style instructions, or refuse unsafe requests — prompt it with the start of a sentence or document.
• At 125M parameters and ~2.2B training tokens, factual accuracy, coherence over long spans, and reasoning are limited. Treat all output as unverified autocomplete, not legal, financial, or professional advice.
• Training data skews heavily toward US case law and SEC filings, so the model's style and any biases reflect that domain (and the OCR noise present in the scanned case-law source).
• License is not specified for this model; the underlying training datasets each carry their own terms — check the dataset cards linked above before redistributing outputs.

Training & Serving Infrastructure

Built end-to-end on Modal: data cleaning, deduplication, tokenizer training, and pretraining all ran as fanned-out CPU/GPU Modal functions against a shared Modal Volume. Inference is served from a Modal web endpoint with token-streaming generation, behind a Next.js frontend deployed on Vercel (link above).