Phi-2-psy

Phi-2-psy is a merge of the following models:

• rhysjones/phi-2-orange
• cognitivecomputations/dolphin-2_6-phi-2

🏆 Evaluation

The evaluation was performed using LLM AutoEval on Nous suite.

Model	AGIEval	GPT4All	TruthfulQA	Bigbench	Average
phi-2-psy	34.4	71.4	48.2	38.1	48.02
phixtral-2x2_8	34.1	70.4	48.8	37.8	47.78
dolphin-2_6-phi-2	33.1	69.9	47.4	37.2	46.89
phi-2-orange	33.4	71.3	49.9	37.3	47.97
phi-2	28.0	70.8	44.4	35.2	44.61

🧩 Configuration

slices:
  - sources:
      - model: rhysjones/phi-2-orange
        layer_range: [0, 32]
      - model: cognitivecomputations/dolphin-2_6-phi-2
        layer_range: [0, 32]
merge_method: slerp
base_model: rhysjones/phi-2-orange
parameters:
  t:
    - filter: self_attn
      value: [0, 0.5, 0.3, 0.7, 1]
    - filter: mlp
      value: [1, 0.5, 0.7, 0.3, 0]
    - value: 0.5
dtype: bfloat16

💻 Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
torch.set_default_device("cuda")
model = AutoModelForCausalLM.from_pretrained("vince62s/phi-2-psy", torch_dtype="auto", trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained("vince62s/phi-2-psy", trust_remote_code=True)
inputs = tokenizer('''def print_prime(n):
   """
   Print all primes between 1 and n
   """''', return_tensors="pt", return_attention_mask=False)
outputs = model.generate(**inputs, max_length=200)
text = tokenizer.batch_decode(outputs)[0]
print(text)

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	62.80
AI2 Reasoning Challenge (25-Shot)	60.84
HellaSwag (10-Shot)	75.52
MMLU (5-Shot)	57.57
TruthfulQA (0-shot)	48.22
Winogrande (5-shot)	75.45
GSM8k (5-shot)	59.21