phi-1.5
1.3B general model trained on 30B tokens of textbook-style synthetic text; near state of the art among sub-10B models at release.
Extended phi-1 from code to common-sense and language reasoning using new synthetic NLP text; released as a base (not instruction-tuned) model for safety research, under MIT licence per its model card.
- Date
- Monday, 11 September 2023
- Lab
- Microsoft
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Training | 150B tokens on a 30B-token dataset 1.3B parameters | company |
Base model, not instruction-tuned or RLHF'd; released for safety research. MIT per the current model card.
Sources
This record was checked against its sources on 6 October 2026. How we check