slm-125m-base/README.md

license, language, library_name, pipeline_tag, tags, datasets

license	language	library_name	pipeline_tag	tags	datasets
apache-2.0	en	transformers	text-generation	llama legal finance small-language-model pretrained-from-scratch causal-lm	HFforLegal/case-law PleIAs/SEC HuggingFaceFW/fineweb-edu

SLM-125M-base

A 125.8M-parameter, Llama-style language model pretrained from scratch on a cleaned, deduplicated, and decontaminated legal + financial corpus. Give it the start of a sentence and it continues in the legal/financial register.

This is a base completer, not a chat model. It was trained on next-token prediction only, so it continues text rather than answering questions.

Model details

Parameters	125,847,552 (~125.8M)
Architecture	Llama-style (LlamaForCausalLM)
Layers	12
Hidden size	768
Attention heads	12 (head dim 64), MHA (12 KV heads)
MLP	SwiGLU, inner 3072
Normalization	RMSNorm, pre-norm
Positional	RoPE (theta 10000)
Context length	1024
Vocab	16,384 (byte-level BPE, trained from scratch)
Embeddings	tied (input = output)
Precision	bf16

Training data

A ~2.04B-token corpus built from three public, ungated sources, streamed and filtered through a deterministic cleaning + dedup + decontamination pipeline:

Source	Content	Share
HFforLegal/case-law	US court opinions	~40%
PleIAs/SEC	SEC filings (10-K, etc.)	~40%
HuggingFaceFW/fineweb-edu	Educational web text	~20%

The corpus was decontaminated against CaseHOLD / LexGLUE (13-gram overlap removed) and deduplicated (exact + MinHash/LSH near-duplicate removal).

Training recipe

Knob	Value
Objective	next-token cross-entropy
Hardware	8× H100 (single-node DDP)
Epochs	1 (~2.04B tokens seen)
Optimizer	AdamW, betas (0.9, 0.95), weight decay 0.1
LR	6e-4 → 6e-5, cosine decay, 200M-token warmup
Global batch	~524,288 tokens/step
Grad clip	1.0
Steps	3,889

Results

Held-out validation perplexity: 11.01 (val loss 2.42) on a 1% held-out split (20.6M tokens). Perplexity was still descending at the end of the single epoch (step 2000 → 12.54, step 3000 → 11.27, final → 11.01), so additional epochs would lower it further. Total compute cost to train: **$11**.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("Sudhanshu1985/slm-125m-base")
model = AutoModelForCausalLM.from_pretrained("Sudhanshu1985/slm-125m-base")

prompt = "The plaintiff alleges that the defendant"
inputs = tok(prompt, return_tensors="pt")
out = model.generate(
    **inputs,
    max_new_tokens=80,
    min_new_tokens=40,
    do_sample=True,
    temperature=0.8,
    top_p=0.95,
    top_k=40,
    repetition_penalty=1.3,
)
print(tok.decode(out[0], skip_special_tokens=True))

Example completions

• "The plaintiff alleges that the defendant" → "breached its contract of employment... In order to prove fraud, a party must show: (1) The existence of a confidential relationship; (2) the absence or violation by one party of any duty owed..."
• "Pursuant to the terms of this Agreement," → "the parties have agreed as follows: 1. ...all disputes... shall be subject to the jurisdiction of said Court..."
• "The Company's net revenues for the fiscal year" → "ended September 30, 1996 were $305.1 million. The operating loss in fiscal 1995 was primarily attributable to..."

Limitations

• Base completer, not a chatbot. It continues text; it does not follow instructions or answer questions.
• Does not know facts. At 125M parameters a model holds only a small amount of usable knowledge; grounded facts would require retrieval (RAG).
• Domain-biased. It speaks the legal register fluently; non-legal prompts drift toward that register.
• Trained on US legal/financial + educational web text; may reflect biases in those sources. Not legal or financial advice.

Provenance

Pretrained from random weights (nothing fine-tuned). Pipeline: stream 3 datasets → rule-based cleaning → dedup + decontaminate → 16K byte-level BPE → pack 1024-token windows → pretrain on 8× H100. Based on the Vizuara AI Labs slm-125m-from-scratch recipe.