slm-125m-base

tirumalaseti/slm-125m-base

This is a 125M-parameter base completion model trained from scratch on a legal/financial-heavy corpus. It is not a chatbot: it was optimized for next-token prediction and works best when prompted with the opening of a sentence or paragraph to continue.

Model summary

• Trainable parameters: 125,847,552 (~125.848M)
• Architecture: 12-layer Llama-style decoder, 768 hidden size, 12 attention heads
• Context length: 1,024 tokens
• Tokenizer: 16,384-token byte-level BPE
• Training tokens seen: 2,500,329,472
• Optimizer steps: 4,769
• Completed epochs over the packed train set: 1.23
• Final validation loss: 2.3035
• Final validation perplexity: 10.01
• Reported spend: $0.00

Training corpus

The packed corpus used for pretraining contains 2,059,674,624 total tokens:

• Train: 2,039,072,768
• Validation: 20,601,856

Realized source mix:

• US case law: 722,081,792 tokens (35.1%)
• SEC filings: 868,714,496 tokens (42.2%)
• Educational web text: 468,878,336 tokens (22.8%)

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "tirumalaseti/slm-125m-base"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id)

prompt = "The plaintiff respectfully moves this Court to"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=96, temperature=0.8, top_p=0.95)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Notes

• This is a base model, not an instruction-tuned assistant.
• It is strongest at continuing legal/financial prose in-register.
• The spend figure comes from the latest visible Modal billing report; Modal billing report may lag; this value reflects the latest locally visible report..