lvesucces/vlm-safety-inspector-dataset

Safety Inspector V2 LoRA Training Dataset & Hyperparameter Specification

This dataset repository contains the offline warm-up SFT dataset and standardized LoRA training configuration for training the Vision-Language Model (VLM) Safety Inspector on the 50 tabletop manipulation scenes (Split into 45 Train + 5 Validation).

---

1. Dataset Overview

• Source Scenes: 50 Tabletop Scenes (45 Train, 5 Val, 0 Test)
• Task Levels: single_step (1 primitive), safety_two_step (2 primitives), multi_step_cleanup (≥3 primitives)
• Unweighted Raw Dataset: 1,300 samples (1,170 train, 130 val)
• Weighted Training Dataset (inspector_v2_train_weighted.jsonl): 2,115 samples
  • PASS samples: 1,080 (8× oversampled to balance class distribution)
  • FAIL samples: 1,035 (Synthetic failures: wrong target, wrong source, missing safety step, non-primitive code, bad route, missing rules, etc.)
• Validation Dataset (inspector_v2_val.jsonl): 130 samples

---

2. Hyperparameter & Protocol Specifications

Hyperparameter	Value	Description
Epochs	2.0	Total passes over the weighted training set
Learning Rate	2e-4	Peak LR with cosine/linear schedule
Per-Device Batch Size	1	Micro-batch size per GPU
Gradient Accumulation	8	Accumulated steps per optimizer step
Effective Batch Size	8	1 * 8 = 8
Total Training Steps	~528	(2115 / 8) * 2 = 528 steps (~264 steps/epoch)
LoRA Rank ($r$)	16	LoRA rank dimension
LoRA Alpha ($\alpha$)	32	LoRA scaling factor
LoRA Dropout	0.05	Dropout probability for LoRA layers
Target Modules	q_proj, v_proj	(Or --target-modules all-linear for full linear projection tuning)
Precision	bfloat16	Mixed precision (or float16 fallback)
Vision Resolution	56x56 min, 512x512 max	Pixel range for visual encoder
Save Steps	100	Save checkpoint every 100 steps
Eval Steps	50	Run validation every 50 steps

---

3. Dataset Directory Structure

• inspector_v2_train_weighted.jsonl: Primary weighted LoRA training dataset (2,115 records)
• inspector_v2_val.jsonl: Validation set (130 records)
• inspector_v2_all.jsonl: Combined full dataset (1,300 unweighted records)
• inspector_v2_summary.json: Detailed breakdown of fail types and split metrics
• lora_training_config.json: Hyperparameter specification in machine-readable JSON format
• train_qwen3vl_lora_inspector.py: Cloud-ready PyTorch/Transformers/PEFT SFT training script

---

4. Quick Start: Cloud Training Command

# Single GPU training
python train_qwen3vl_lora_inspector.py \
    --model Qwen/Qwen2-VL-7B-Instruct \
    --train-jsonl inspector_v2_train_weighted.jsonl \
    --val-jsonl inspector_v2_val.jsonl \
    --epochs 2.0 \
    --learning-rate 2e-4 \
    --batch-size 1 \
    --grad-accum 8 \
    --lora-r 16 \
    --lora-alpha 32 \
    --output-dir ./outputs/qwen3vl_inspector_lora

# Multi-GPU Distributed (DDP / DeepSpeed)
torchrun --nproc_per_node=4 train_qwen3vl_lora_inspector.py \
    --model Qwen/Qwen2-VL-7B-Instruct \
    --train-jsonl inspector_v2_train_weighted.jsonl \
    --val-jsonl inspector_v2_val.jsonl \
    --epochs 2.0 \
    --learning-rate 2e-4 \
    --batch-size 1 \
    --grad-accum 2 \
    --lora-r 16 \
    --lora-alpha 32 \
    --target-modules all-linear \
    --output-dir ./outputs/qwen3vl_inspector_lora