Atlas

Models

← All models

Llama-3.1-Tülu-3-8B

Allen Institute for AI (Ai2) · 2024-11-21 · 8B parameters

Llama-3.1-Tulu-3-8B is Ai2's post-trained 8B Llama 3.1 model. It was produced with the fully open Tülu 3 recipe, combining supervised fine-tuning, Direct Preference Optimization, and reinforcement learning with verifiable rewards.

Benchmark scores

BenchmarkScore
Arena Text — Business, Management & Financial Ops (No Style Control)1196
Arena Text — Chinese (No Style Control)1175
Arena Text — Creative Writing (No Style Control)1182
Arena Text — English (No Style Control)1215
Arena Text — Entertainment, Sports & Media (No Style Control)1161
Arena Text — Exclude Ties (No Style Control)1060
Arena Text — Exclude Ties (Style Controlled)1101
Arena Text — Hard Prompts (No Style Control)1175
Arena Text — Hard Prompts English (No Style Control)1180
Arena Text — Instruction Following (No Style Control)1175
Arena Text — Life, Physical & Social Science (No Style Control)1194
Arena Text — Longer Query (No Style Control)1181
Arena Text — Multi-turn (No Style Control)1154
Arena Text — Non-English (No Style Control)1169
Arena Text — Non-English (Style Controlled)1199
Arena Text — Occupational Mathematical (No Style Control)1193
Arena Text — Russian (No Style Control)1192
Arena Text — Software & IT Services (No Style Control)1192
Arena Text — Writing, Literature & Language (No Style Control)1178
BBH (Open LLM Leaderboard v2)40.5
Capability108.2
GPQA (Open LLM Leaderboard v2)29.9
IFEval (Open LLM Leaderboard v2)82.7
LMArena Text Arena Chinese Elo1220
LMArena Text Arena Coding Elo1253
LMArena Text Arena Creative Writing Elo1197
LMArena Text Arena Elo1221
LMArena Text Arena English Elo1239
LMArena Text Arena Hard Prompts Elo1223
LMArena Text Arena Hard Prompts English Elo1225
Loading Atlas data…