AI Potluck
Model components / Training & synthetic datasets

OpenHermes 2.5

Teknium

OpenHermes 2.5 is a large synthetic instruction-following dataset of ~1M conversational examples compiled and generated primarily from GPT-4 outputs across math, code, roleplay, and general-knowledge tasks. Created by Teknium as an expansion of OpenHermes 1, it was used to train the popular OpenHermes 2.5 fine-tunes and is widely used as a base SFT mix for open instruction-tuned models.

Verified live 2026-06-22 via primary sources. Publicly downloadable and ungated with a full card, though an explicit license string is not stated on the card.

Openness

5 medium confidence
5.0
license
not-clearly-stated-on-card(commonly cited as permissive)
card
present
ungated
yes
format
JSON/Parquet
rows
~1M

Publicly downloadable and ungated with a full card, though an explicit license string is not stated on the card.

Adoption

3 high confidence
3.0

14,982 monthly downloads on Hugging Face, graded on the training-corpus bands.

Capability

3 high confidence
3.0

Attributed gains inside SmolTalk (MMLU/BBH/WinoGrande) but no standalone ablation; superseded by Hermes 3 and survives as a 100k SmolTalk slice.

Unchanged since 2026-07-04 (last edited, not re-checked)