Last updated: 2026-09-04
FINEFORGE
Fine-tuning datasets, forged from video & papers.
The expensive step of a fine-tune is not training: it is gathering the examples. Almost every tool in the space assumes you already have the documents. FineForge takes care of the step before — finding them and deciding which ones are worth it.
In production, with real users: fineforgeai.com
What it does
You describe the domain in plain language. FineForge searches YouTube, arXiv and OpenAlex, a model scores every candidate from 0 to 100 against what you asked for, you approve or discard them one by one, and what survives comes out as training-ready JSONL plus one PDF per source.
| Stage | What happens | What you control |
|---|---|---|
| Search | YouTube, arXiv and OpenAlex from a single prompt | The prompt and the sources |
| Scoring | Every candidate gets a 0–100 relevance score | The threshold |
| Curation | You approve or reject each one by hand | Everything |
| Export | JSONL ready to train, PDF summaries per source | The format |
Who it is for
Teams that fine-tune models on a specific domain and are spending days digging through videos and papers by hand. Solo researchers with the same problem.
What it is not
- Not a training service: it builds the dataset, you train wherever you want.
- Not a scraper that dumps everything: the scoring and the manual curation are the point.
- Not a general search engine: it is built around the question "is this a good training example?"
Facts you can check
| Status | In production, with real users |
| Sources | YouTube, arXiv, OpenAlex |
| Output | JSONL plus PDF summaries |
| URL | fineforgeai.com |
| Built by | Forge-AI, 2026 |