Merge pull request #241 from LLaVA-VL/ov-chat-doc

add dpo training script to Ov chat doc

Author: ChunyuanLI

docs/LLaVA_OneVision_Chat.md

@@ -94,13 +94,29 @@ Using the feedback data obtained in `Step 2`, we conduct DPO training in an iter
 
 This iterative process is repeated for `N=3` rounds in total, with each round refining the model’s ability to generate high-quality visual chat responses by progressively incorporating feedback from both human and AI assessments.
 
+**Training script and data format**
+
+- Example training script: [`/scripts/train/dpo_ov7b.sh`](../scripts/train/dpo_ov7b.sh)
+- Format of training data:
+~~~json
+{
+  "id": "<image-id>",
+  "image": "<image path under args.image_folder>",
+  "prompt": "<input prompt/question>",
+  "chosen": "<chosen model response>",
+  "rejected": "<rejected model response>"
+}
+~~~
+
+
 ------
 
 Stay tuned on how we develop AI feedback for self-improvement LMMs!
 
 *Contributors to LLaVA-OneVision-Chat: [Tianyi Xiong](https://tyxiong23.github.io/), [Bo Li](https://brianboli.com/), [Dong Guo](https://www.linkedin.com/in/dongguoset/), [Huizhuo Yuan](https://scholar.google.com/citations?user=8foZzX4AAAAJ), [Quanquan Gu](https://web.cs.ucla.edu/~qgu/), [Chunyuan Li](https://scholar.google.com/citations?user=Zd7WmXUAAAAJ)*
 
 
+
 ### Citation
 
 If you find it useful for your research and applications, please cite related papers/blogs using this BibTeX:

scripts/train/dpo_ov7b.sh

@@ -0,0 +1,69 @@
+export OMP_NUM_THREADS=8
+export NCCL_IB_DISABLE=0
+export NCCL_IB_GID_INDEX=3
+# export NCCL_IB_HCA=${ARNOLD_RDMA_DEVICE}
+export NCCL_SOCKET_IFNAME=eth0
+export NCCL_DEBUG=INFO
+
+VISION_MODEL_VERSION="google/siglip-so400m-patch14-384"
+VISION_MODEL_VERSION_CLEAN="${VISION_MODEL_VERSION//\//_}"
+
+# DPO Stage
+PROMPT_VERSION="qwen_1_5"
+SFT_MODEL="lmms-lab/llava-onevision-qwen2-7b-ov"
+EPOCH=1
+beta=0.1
+
+DPO_RUN_NAME="llava-onevision-qwen2-7b-ov_dpo-beta${beta}-epoch${EPOCH}"
+DPO_CLEAN_NAME="${DPO_RUN_NAME##*/}"
+OUTPUT_DIR="<your-output-folder>/${DPO_CLEAN_NAME}"
+DATA_PATH="<your-data-path>"
+
+echo $DPO_RUN_NAME
+
+ACCELERATE_CPU_AFFINITY=1 torchrun --nproc_per_node="${NUM_GPUS}" --nnodes="${NNODES}" --node_rank="${RANK}" --master_addr="${ADDR}" --master_port="${PORT}" \
+    llava/train/train_dpo.py \
+    --deepspeed scripts/zero3.json \
+    --model_name_or_path=${SFT_MODEL} \
+    --dpo_alpha=1.0 \
+    --beta=${beta} \
+    --gamma=0 \
+    --version $PROMPT_VERSION \
+    --data_path=$DATA_PATH \
+    --image_folder "<your-image-folder>" \
+    --mm_tunable_parts="mm_vision_tower,mm_mlp_adapter,mm_language_model" \
+    --unfreeze_mm_vision_tower True \
+    --vision_tower ${VISION_MODEL_VERSION} \
+    --mm_projector_type mlp2x_gelu \
+    --mm_vision_select_layer -2 \
+    --mm_use_im_start_end False \
+    --mm_use_im_patch_token False \
+    --group_by_modality_length True \
+    --image_aspect_ratio anyres_max_9 \
+    --image_grid_pinpoints "(1x1),...,(6x6)" \
+    --mm_patch_merge_type spatial_unpad \
+    --bf16 True \
+    --run_name $DPO_CLEAN_NAME \
+    --output_dir $OUTPUT_DIR \
+    --num_train_epochs $EPOCH \
+    --per_device_train_batch_size 1 \
+    --per_device_eval_batch_size 1 \
+    --gradient_accumulation_steps 8 \
+    --evaluation_strategy "no" \
+    --save_strategy "steps" \
+    --save_steps 1000 \
+    --save_total_limit 1 \
+    --learning_rate 5e-7 \
+    --weight_decay 0. \
+    --warmup_ratio 0.1 \
+    --lr_scheduler_type "cosine" \
+    --logging_steps 1 \
+    --tf32 True \
+    --model_max_length 32768 \
+    --gradient_checkpointing True \
+    --dataloader_num_workers 4 \
+    --lazy_preprocess True \
+    --report_to wandb \
+    --dataloader_drop_last True
+
+