A 125-million-parameter model, pretrained from random weights on legal and financial text, then fine-tuned to answer questions from a passage you give it. Paste a passage, ask about it, and it answers using only what's there.
The base model only continues text — it can't answer questions. This is that base model after supervised fine-tuning on 13,774 teacher-generated, LLM-judged Q&A pairs, each grounded in a source passage. It now reads a passage and answers about it.
What it does well: extractive questions — who, what, where, which claims, what did the court hold, what does a clause require. It pulls the answer from the passage and phrases it directly.
Honest limits: at 125M it cannot do arithmetic (“6% of $50,000”, revenue growth) and has almost no world knowledge of its own — take away the passage and it has little to say. It answers from what you give it, not from memory.
Passages drawn from US case law, SEC filings, and educational web text. First call may take ~10–20s while the model wakes from idle.
The whole thing — a language model built from random weights to a working grounded-Q&A assistant — cost about the price of a large pizza. Every figure below is measured spend, not an estimate.
| Pretraining data pipeline + 8×H100, 4 epochs, 2.04B tokens + sampling |
$14.43 |
| SFT dataset 13,774 Q&A pairs — GPT‑5.4‑nano generates, GPT‑5.4‑mini judges, + pilots |
$11.50 |
| Fine-tuning 1×H100, 2 epochs kept, 3.9 minutes wall-clock |
$0.21 |
| Data pipeline compute chunk / dedup / tokenize on Modal CPU |
$0.60 |
| Serving scale-to-zero endpoint — $0 while idle |
$0.00 |
| Total | ~$27 |
The output side dominated: at this scale, the teacher LLM writing and grading the training data ($11.50) cost nearly as much as pretraining the model itself ($14.43). Fine-tuning — the step everyone pictures as “the expensive part” — was 21 cents. The judge alone was 60% of the dataset bill, because verifying that an answer is truly grounded is harder than writing it.