Atlas

Models

← All models

Neural Chat 7B v3-1

Intel · 2023-11-14 · 7.2B parameters

Neural Chat 7B v3-1 is Intel's instruction-following fine-tune of Mistral-7B-v0.1, fine-tuned with supervised learning on Open-Orca/SlimOrca and aligned with Direct Preference Optimization on Intel/orca_dpo_pairs using eight Intel Gaudi 2 accelerators. Intel documents an 8,192-token context for this release, while the checkpoint configuration records max_position_embeddings of 32,768 and a 4,096-token sliding-attention window.

Benchmark scores

BenchmarkScore
BBH (Open LLM Leaderboard v2)50.5
Capability103.5
EQ-Bench v264.8
EQ-Bench v2 + MAGI-Hard Combined50.7
GPQA (Open LLM Leaderboard v2)29.0
IFEval (Open LLM Leaderboard v2)46.9
MAGI-Hard36.6
MATH Level 5 (Open LLM Leaderboard v2)3.5
MMLU-Pro (Open LLM Leaderboard v2)26.8
MuSR (Open LLM Leaderboard v2)49.8
Loading Atlas data…