maldv/badger-l3-instruct-32k

updated with fixed tokenizer config

Badger/δ Llama 3 Instruct 32k

I haven't been releasing my base merges so far, but this one seems worthy.

Badger is a recursive maximally disjoint pairwise normalized fourier interpolation of the following models:

models = [
 'Einstein-v6.1-Llama3-8B',
 'L3-TheSpice-8b-v0.8.3',
 'dolphin-2.9-llama3-8b',
 'Configurable-Hermes-2-Pro-Llama-3-8B',
 'MAmmoTH2-8B-Plus',
 'Pantheon-RP-1.0-8b-Llama-3',
 'Tiamat-8b-1.2-Llama-3-DPO',
 'Buzz-8b-Large-v0.5',
 'Kei_Llama3_8B',
 'Llama-3-Lumimaid-8B-v0.1',
 'llama-3-cat-8b-instruct-pytorch',
 'Llama-3SOME-8B-v1',
 'Roleplay-Llama-3-8B',
 'Llama-3-LewdPlay-8B-evo',
 'opus-v1.2-llama-3-8b-instruct-run3.5-epoch2.5',
 'meta-llama-3-8b-instruct-hf-ortho-baukit-5fail-3000total-bf16',
 'Poppy_Porpoise-0.72-L3-8B',
 'Llama-3-8B-Instruct-norefusal',
 'Meta-Llama-3-8B-Instruct-DPO',
 'badger',
 'Llama-3-Refueled',
 'Llama-3-8B-Instruct-DPO-v0.4',
 'Llama-3-8B-Instruct-Gradient-1048k',
 'Mahou-1.0-llama3-8B',
 'Llama-3-SauerkrautLM-8b-Instruct',
 'Llama-3-Soliloquy-8B-v2'
]

I have included the notebook code I used to generate the model, for any that are curious. I have adjusted the config for rope scale 4, and 16k-32k context both seem coherent.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	69.49
AI2 Reasoning Challenge (25-Shot)	63.65
HellaSwag (10-Shot)	81.40
MMLU (5-Shot)	67.13
TruthfulQA (0-shot)	55.02
Winogrande (5-shot)	77.35
GSM8k (5-shot)	72.40