Nous Capybara 7B V1
Nous Research · 2023-09-20 · 6.7B parameters
Nous Capybara 7B V1 is Nous Research's Llama-2-based 7B conversational fine-tune, trained for multiple epochs on roughly 20,000 carefully curated conversational examples. Its data was created with Nous's Amplify-Instruct synthesis method, which combines seed distributions and techniques associated with Airoboros, Evol-Instruct, Orca, Vicuna, and others; the model card says most fine-tuning tokens were newly synthesized in house and that more than 10,000 examples were multi-turn.
Benchmark scores
| Benchmark | Score |
|---|---|
| EQ-Bench v2 | 34.4 |
| EQ-Bench v2 + MAGI-Hard Combined | 32.3 |
| MAGI-Hard | 30.2 |