Vizuara AI Labs · trained from scratch

SLM‑125M

A 125-million-parameter Llama-style language model, pretrained from random weights on a legal and financial corpus. Give it the start of a sentence and it continues in the legal register.

125.8M
parameters
16,384
vocab
1,024
context
2.18B
train tokens
10.87
val perplexity
1
epoch
your prompt (a prefix to continue)
ready
The completion will appear here.

What this is base completer

This is a base model, not a chatbot. It was trained on next-token prediction only, so it continues text rather than answering questions. Prompt it with the opening of a sentence and watch it complete the thought.

The honest quality metric is held-out validation perplexity: 10.87 (lower is better), measured over a 1% held-out split (~22M tokens). It speaks the legal register (case-citation phrasing, procedural language) but it does not know facts — at 125M parameters a model holds only about 31MB of usable knowledge. Grounded facts would need retrieval (RAG).

model & training spec
architectureLlama-style decoder (RoPE, SwiGLU, RMSNorm, MHA)
layers / hidden / heads12 / 768 / 12 (head dim 64), kv-heads 12
parameters125,847,552 (tied embeddings)
vocab / context16,384 byte-level BPE / 1,024 tokens
train / val tokens2.18B / 22.0M (99/1 split)
corpus mixUS case law ~33% · SEC filings ~39% · edu web ~28%
optimizerAdamW (0.9, 0.95), wd 0.1, grad-clip 1.0
lr schedule6e-4 → 6e-5 cosine, 200M-token warmup
batch / precision524,288 tokens/step, bf16
compute8×H100 DDP · 1 epoch · ~24 min
final val perplexity10.87
stream 3 datasets clean rule chain dedup + decontaminate 16K byte-level BPE pack 1024-token windows pretrain 8×H100, 1 epoch

Corpus: US case law, SEC filings, educational web text. First call may take ~15–30s while the CPU endpoint wakes from idle.