MPP-Qwen-Next_ckpt-and-data

• pretrain: including model.pth, the linear projection weight after the MPP-Qwen-next Pretrain Stage.
• llava_instruct: process script and annotations(json file)
• llava_pretrain: annotations(json file)
• videochatgpt: process script and annotations(json file)

MPP-Qwen14B pretrain/sft data and pretrain ckpt(linear projection weight)

pretrain:

• model.pth: the linear projection weight after the MPP-Qwen14B Pretrain Stage.

llava_instruction_100k:

• complex_reasoning_77k and details_23k of llava_instruction_tuning data. I've converted it into MPP-Qwen14B format.

llava_pretrain_558k:

• 558K pretrain data of LLaVA. I've converted it into MPP_Qwen14B format.

instruction-data, checkpoint and logs(also 14B model)

Release the instruction data to align MiniGPT4 to Qwen-Chat LLM model and my checkpoint(all 10 epochs trained by lavis/projects/instruction_tuning/train.yaml).

Release MiniGPT4Qwen14B model checkpoint and train logs(20 epochs trained by lavis/projects/pp_qwen14b/train_pp.yaml). The files are compressed in pp_14b_ckpt-logs.zip.