Fine tuning data that fits your model.
Generic instruction data gives you a generic model. We build fine tuning datasets around your domain and your use case, sourced from the websites and materials you specify.
Datasets we build
Whether you are doing supervised fine tuning on a niche domain or aligning a model with preference data, the raw material matters more than the training script. We supply that raw material.
- Instruction and SFT data. Prompt and response pairs built from real source material in your domain, formatted for your training framework.
- Domain corpora. Large, clean text collections from the sites and document types you care about, deduplicated and filtered.
- Preference pairs. Chosen and rejected response sets for DPO or RLHF style training, written against your rubric.
- Evaluation sets. Held out test data that mirrors your production use case, so your benchmarks mean something.
Cleaning and decontamination
Raw web text is full of boilerplate, duplicates, and near duplicates. Every dataset goes through cleaning passes before delivery, and we check against your benchmark sets on request so evaluation results stay trustworthy.
Private by default
Your fine tuning data is a competitive asset. Engagements are private, payment is in crypto, and nothing about your project is shared or reused.
Common questions
Which formats do you deliver in?
Whatever your trainer expects. JSONL with your field names is most common, but we also deliver plain text, Parquet, or chat templated formats.
Can you build data for a very specific domain?
That is the normal case for us. The more specific the domain, the more a custom dataset beats anything you can download.
Do you offer preference data as well as SFT data?
Yes. See our RLHF data page, or just describe what you need in an email to data@nexorpus.com.
Describe the data. We deliver it.
One email is all it takes. We reply with questions, samples, and a quote that beats the market.
data@nexorpus.com