Confused embed_dim across different implementation

[tsWen0309]

Hi, I want to post-training a model and test on the MTEB benchmark.
During implementation , I found that load pre-trained model in different way results in different embed_dim.
For example:
model = SentenceTransformer("dunzhang/stella_en_1.5B_v5") results in a embedding output dim =1024, exact as mentioned in https://huggingface.co/dunzhang/stella_en_1.5B_v5.
while using model = mteb.get_model('dunzhang/stella_en_1.5B_v5') rsults in a embedding output dim = 1536.
The difference in embedding dim make me so confused and also the test results also differs due to the embedding dim.

Here's the code I use:

import mteb
from sentence_transformers import SentenceTransformer
from transformers import AutoTokenizer, AutoModel
from typing import Dict, List, Optional, Union
from mteb.encoder_interface import PromptType
from mteb.models.wrapper import Wrapper
import numpy as np

class my_model(Wrapper):
def init(self,model_name):
super().init()
# self.model = mteb.get_model(model_name)
self.model = SentenceTransformer(model_name)
def encode(
self,
sentences: list[str],
task_name: str,
prompt_type: PromptType | None = None,
**kwargs,
) -> np.ndarray:
embeddings = self.model.encode(sentences, task_name=task_name)
# print(embeddings.shape)
return embeddings

tasks_name: Dict[str, list] = {
'classification':[
"Banking77Classification",
],
'retrieval':[
'ClimateFEVER',
'DBPedia',
'FEVER',
'QuoraRetrieval',
'SciFact',
'TRECCOVID'
]
}

model = my_model("dunzhang/stella_en_1.5B_v5")

tasks = mteb.get_tasks(
tasks=tasks_name['classification'],
languages=['eng'],
)
evaluation = mteb.MTEB(tasks=tasks)
results = evaluation.run(
model,
eval_splits=["test"],
output_folder="results",
)

Also , another problem is. Under the implementation of MTEB, is the MRL of this model(https://huggingface.co/dunzhang/stella_en_1.5B_v5) also implemented? If so, how can I use it? If not, how to reproduce the MTEB results duo to SentenceTransformer(cause the MRL representation is implemented in SentenceTransformer)?

[Samoed]

Stella is an instruction model and uses this InstructLoader. What does MRL stand for? Also, for the future, could you format code with ``` for better readability?

[tsWen0309]

Stella is an instruction model and uses this InstructLoader. What does MRL stand for? Also, for the future, could you format code with ``` for better readability?

Thanks for your reply. I apologize for the inconvient. I will use the InstructLoader to see the results. MRL stands for Matryoshka Representation Learning.

[KennethEnevoldsen]

This seems to be an usage question, will move it into discussion but feel free to continue the discussion there