Datasets:
Additional information detailing the creation of the dataset
Thank you for the work on the No Robots dataset. Will a technical report with more information on the prompt creation process follow? For a low resource language I would like to replicate the annotation procedure and am for this reason curious how the prompts were selected and created, which quality control measures were in place and what tooling has been used to perform the annotations.
Hi @davodavo ! While the No Robots paper and model card document the general procedure, I can share what worked for me when building a similar human-annotated synthetic instruction dataset (and what I'd recommend for a low-resource language replication):
Prompt selection
The No Robots approach uses stratified category sampling β the key insight is to define your categories first, then constrain annotators to them, rather than letting them free-write and categorizing post-hoc. For a low-resource language, I'd recommend:
- Pick 6-8 task categories (e.g. Generation, Rewrite, Summarize, QA, Classification, Brainstorming, Chat, Code) that match realistic use cases for your language.
- Pre-assign each annotator a target distribution per session: "write 3 Generation, 2 QA, 1 Code prompts this session." This avoids the natural human bias toward whichever category is easiest.
- Seed annotators with 5-10 examples per category before they start β reduces ambiguity about what each category means.
Quality control
For the No Robots format, the two biggest quality signals are:
- Prompt naturalness: Does the instruction sound like something a real user would type? Have a second annotator rate on a 1-5 scale; keep β₯4.
- Response groundedness: Does the response actually answer the specific prompt (not a generic similar one)? Spot-check 10% manually, or run a cross-check where annotators review each other's pairs.
A lightweight pipeline: annotate in Argilla (free, self-hostable), export to JSONL, run deduplication by fuzzy prompt similarity (MinHash or sentence-transformer embeddings), then a final human review pass on flagged near-duplicates.
Tooling
- Annotation: Argilla is well-suited to instruction-response pairs. Label Studio also works but requires more schema setup.
- Quality filtering:
sentence-transformersfor embedding-based near-dedup;langdetectto catch annotator code-switching into another language. - Format: Match the No Robots schema exactly (
{"prompt": ..., "category": ..., "messages": [{"role": "user", "content": ...}, {"role": "assistant", "content": ...}]}) β this makes it compatible with SFT trainers like TRL out of the box.
For low-resource languages specifically
Budget 3-4Γ more time per example vs. English β annotators will hit missing vocabulary and awkward phrasings more often. I'd suggest a two-stage review: annotator writes, then a native speaker reviewer polishes before the pair goes into the dataset. Even 500 high-quality native-reviewed examples is more valuable than 2,000 machine-translated ones.
Hope that helps β good luck with the replication!