Nemotron-4 340B Instruct
NVIDIA · 2024-06-14
Nemotron-4 340B Instruct is NVIDIA's open-access chat model aligned from Nemotron-4 340B Base and released on June 14, 2024, intended mainly for English single- and multi-turn chat and synthetic-data generation. It went through code and general supervised fine-tuning, DPO and three rounds of Reward-aware Preference Optimization, with over 98% of the alignment data synthetically generated, and keeps the base model's 96-layer GQA architecture (about 341 billion parameters, no exact checkpoint count published) and 4,096-token context. It is distributed in NeMo format under the NVIDIA Open Model License.
Benchmark scores
| Benchmark | Score |
|---|---|
| AlpacaEval 2.0 (length-controlled win rate) | 41.5 |
| Arena-Hard-Auto v0.1 | 54.2 |
| Capability | 118.1 |
| GSM8K | 92.3 |
| HumanEval | 73.2 |
| IFEval | 79.9 |
| IFEval (instruction-level strict accuracy) | 86.1 |
| MBPP | 75.4 |
| MMLU | 78.7 |
| MT-Bench (GPT-4-Turbo judge, corrected references) | 8.2 |