Implementation:OpenGVLab InternVL Correctness Build Data

Knowledge Sources	InternVL
Domains	Alignment, Data_Engineering
Last Updated	2026-02-07 00:00 GMT

Overview

Concrete tool for constructing correctness-based preference data via shell script pipelines provided by the InternVL MPO data construction framework.

Description

The correctness_build_data.sh script orchestrates the preference data construction pipeline. It runs mmpr_data_pipeline_correctness_postprocess.py across multiple tile configurations to generate MMPR-format preference pairs from model-generated reasoning samples.

A companion visualprm_build_data.sh constructs VisualPRM data using Monte Carlo step-level reward estimation.

Usage

Run these scripts before MPO training to generate the preference dataset. They require pre-generated model sampling outputs.

Code Reference

Source Location

• Repository: InternVL
• File: internvl_chat/shell/internvl3.0/mpo_data_construction/correctness_build_data.sh
• Lines: L1-34

Signature

# correctness_build_data.sh
# Iterates over tile configurations and runs postprocessing

for MAX_NUM in 1 6 12 18 24; do
    python tools/reasoning_data_pipeline/mmpr_data_pipeline_correctness_postprocess.py \
        --num-pairs-per-key 15 \
        --max-lines 1200000 \
        --answer-fix \
        --max-num $MAX_NUM \
        --input-dir $INPUT_DIR \
        --output-dir $OUTPUT_DIR
done

# visualprm_build_data.sh
for MAX_NUM in 1 6 12 18 24; do
    python tools/reasoning_data_pipeline/visualprm_data_pipeline_postprocess.py \
        --mc-threshold 0.0 \
        --max-num $MAX_NUM \
        --input-dir $INPUT_DIR \
        --output-dir $OUTPUT_DIR
done

Import

cd internvl_chat
bash shell/internvl3.0/mpo_data_construction/correctness_build_data.sh
bash shell/internvl3.0/visualprm_data_construction/visualprm_build_data.sh

I/O Contract

Inputs

Name	Type	Required	Description
Model sampling outputs	Directory	Yes	Pre-generated reasoning samples from the model
--num-pairs-per-key	int	No	Max preference pairs per question (default 15)
--max-lines	int	No	Per-tile-config sample limit (default 1200000)
--mc-threshold	float	No	Monte Carlo threshold for VisualPRM (default 0.0)

Outputs

Name	Type	Description
MMPR JSONL	Files	Preference data with question, chosen, rejected fields and associated images

Usage Examples

Build Correctness Preference Data

cd internvl_chat

# Step 1: Generate model samples (prerequisite)
bash shell/internvl3.0/mpo_data_construction/correctness_mmpr_8b.sh

# Step 2: Build preference pairs from samples
bash shell/internvl3.0/mpo_data_construction/correctness_build_data.sh

# Step 3: Build VisualPRM data
bash shell/internvl3.0/visualprm_data_construction/visualprm_build_data.sh

Related Pages

Implements Principle

• Principle:OpenGVLab_InternVL_Preference_Data_Construction

Requires Environment

• Environment:OpenGVLab_InternVL_PyTorch_CUDA