Modelpublic

jsun39/Cosine-Beta-KD-Instance

sourceHugging Faceapache-2.0updated 5mo agoView on Hugging Face

0likes8downloads

Model Card

Cosine-Beta-KD-Instance

A 1.7B multimodal LLM checkpoint distilled with Cosine-KD + Beta-KD (Instance-level uncertainty weighting), built on top of MobileVLM with `MobileLLaMA-1.4B-Chat` as the language backbone.

This checkpoint corresponds to the `Beta-KD (Instance)` row of the model zoo in Beta-KD: Uncertainty-Aware Knowledge Distillation for Multimodal Large Language Models.

Model Details

Item	Value
Architecture	MobileVLM (CLIP visual encoder + LDP projector + MobileLLaMA LLM)
Language model	MobileLLaMA 1.4B
Distillation losses	Cosine-KD (logit alignment) + Beta-KD instance-level uncertainty loss
Training step	`checkpoint-18000`
Total params	~1.7B
Precision	fp16

Evaluation

Evaluated on six standard multimodal benchmarks (no beam search, greedy decoding to match the chat-demo behavior).

Method	LLM	MME<sup>P</sup>	MME<sup>A</sup>	GQA	VQA<sup>T</sup>	POPE	MMB<sup>dev</sup>	SQA<sup>I</sup>	Avg.
Cosine-KD baseline	MobileLLaMA 1.4B	1308.4	65.4	59.9	52.2	84.6	57.1	61.3	63.4
+ Beta-KD (Task)	MobileLLaMA 1.4B	1352.0	67.6	60.8	53.9	85.4	59.1	61.2	64.7
+ Beta-KD (Instance) (this model)	MobileLLaMA 1.4B	1350.3	67.5	61.2	54.2	86.0	60.2	62.9	65.3

Usage

python

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

repo_id = "jsun39/Cosine-Beta-KD-Instance"

tokenizer = AutoTokenizer.from_pretrained(repo_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    repo_id,
    torch_dtype=torch.float16,
    trust_remote_code=True,
).cuda()

For full inference (image + text), please follow the inference example in the Beta-KD repo — the visual encoder / projector loading, image preprocessing, and chat template are described there.

Files

This repo contains only the files needed for inference:

—pytorch_model.bin — fp16 weights
—config.json, generation_config.json
—tokenizer.model, tokenizer_config.json, special_tokens_map.json

DeepSpeed optimizer / RNG / trainer states are intentionally not uploaded.

Citation

bibtex

@article{sun2026betakd,
  title   = {Beta-KD: Uncertainty-Aware Knowledge Distillation for Multimodal
             Large Language Models},
  author  = {Sun, Jingchen and Han, Shaobo and Patel, Deep and Kohno, Wataru and Jin, Can and Chen, Changyou},
  journal = {CVPR},
  year    = {2026}
}

License

Released under the Apache-2.0 license, inheriting from MobileVLM and MobileLLaMA. The visual encoder and any third-party data follow their original licenses.