Skip to content

CLIP+Q-Former+projector+Qwen, and BLIP+projector+Qwen implementation#18

Open
qingwuu wants to merge 38 commits intomainfrom
TrainLoop
Open

CLIP+Q-Former+projector+Qwen, and BLIP+projector+Qwen implementation#18
qingwuu wants to merge 38 commits intomainfrom
TrainLoop

Conversation

@qingwuu
Copy link
Collaborator

@qingwuu qingwuu commented Jan 27, 2026

No description provided.

qingwuu and others added 30 commits November 19, 2025 18:08
HF dataset, "answer" be corrected to json format ; train.py, able to run
train restructure, TODO: debug
fix some errors in training files
metrics: bleu and rouge;
 BARTScore not implemented yet ,need furtherconsideration
Trainer,: compute loss and prediction step override
labels concatenation should not happen in prediction step. now this issue has been fixed
stage name added, more readable
some training arguments
1. T and T_pad
2.promt refine
3.training config
4.dataset refine
Qformer impelmentation, and data strcuture/ dimension matching
Q-former , data dimension alignment
test print, robust
single shot cross attention
lack of FFN, LayerNorm
rearrange files
qingwuu and others added 8 commits January 27, 2026 02:31
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant