phi-1 ('Textbooks Are All You Need')
1.3B code model trained 4 days on 8 A100s on 'textbook quality' web plus synthetic data reaches 50.6% HumanEval.
Launched the Phi thesis that data quality can substitute for scale. The model trained on 6B filtered web tokens plus 1B GPT-3.5-generated tokens beat much larger code models on HumanEval and MBPP (55.5%).
- Date
- Tuesday, 20 June 2023
- Lab
- Microsoft
- Kind
- model
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| HumanEval pass@1 | 50.6% MBPP 55.5%; 1.3B params, 4 days on 8 A100s | company |
arXiv v1 2023-06-20; Wikipedia gives a 2023-06-21 release date. Whether synthetic data leaked benchmark content was debated.
Sources
This record was checked against its sources on 6 October 2026. How we check