mzbac/qwen-1_5-7B-2x3-hf

The Moe model built on top of Qwen1.5-7B-Chat, Qwen1.5-7B and Crystalcareai/CrystalQwen-1.5-7B, Then qlora was applied to all layers of q,v, and gate linear on WizardLM_evol_instruct_70k via mlx. The model was created using a script from https://github.com/mzbac/mlx-moe

Evaluation

Qwen-1_5-2x3-hf

MMLU

Groups	Version	Filter	n-shot	Metric	Value		Stderr
- humanities	N/A	none	0	acc	0.6488	±	0.0237
- other	N/A	none	0	acc	0.6294	±	0.0302
- social_sciences	N/A	none	0	acc	0.6905	±	0.0281
- stem	N/A	none	0	acc	0.5227	±	0.0375

CMMLU

Groups	Version	Filter	n-shot	Metric	Value		Stderr
cmmlu	N/A	none	0	acc	0.6966	±	0.0333
		none	0	acc_norm	0.6966	±	0.0333

GSM8K

Tasks	Version	Filter	n-shot	Metric	Value		Stderr
gsm8k	2	get-answer	5	exact_match	0.4102	±	0.0135

Qwen1.5-7B-Chat

MMLU

Groups	Version	Filter	n-shot	Metric	Value		Stderr
- humanities	N/A	none	0	acc	0.6533	±	0.0239
- other	N/A	none	0	acc	0.6321	±	0.0301
- social_sciences	N/A	none	0	acc	0.6934	±	0.0282
- stem	N/A	none	0	acc	0.5329	±	0.0376

CMMLU

Groups	Version	Filter	n-shot	Metric	Value		Stderr
cmmlu	N/A	none	0	acc	0.6879	±	0.0338
		none	0	acc_norm	0.6879	±	0.0338

GSM8K

Tasks	Version	Filter	n-shot	Metric	Value		Stderr
gsm8k	2	get-answer	5	exact_match	0.0425	±	0.0056

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "mzbac/qwen-1.5-2x3-hf"

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    load_in_4bit=True,
    trust_remote_code=True,
)

tokenizer = AutoTokenizer.from_pretrained(model_id)

chat = [
    {"role": "user", "content": "how backpropagation works?"},
    {"role": "assistant", "content": "\n"},
]

text = tokenizer.apply_chat_template(chat, tokenize=False)

inputs = tokenizer.encode(text, return_tensors="pt").to("cuda")

generate_kwargs = dict(
    input_ids=inputs,
    temperature=0.6,
    max_new_tokens=500,
    do_sample=True,
)

outputs = model.generate(**generate_kwargs)
print(tokenizer.decode(outputs[0]))