bear7011/gemma-4-e4b-vatex4K_FT

gemma-4-e4b-kinetics_4K_VATEX

SFT training dataset for Gemma 4 E4B video action recognition, combining Kinetics-40K and VATEX (overlap-cleaned).

Dataset Contents

File	Entries	Description
train_described.json	40,000	Kinetics-40K — converted from k400 annotation CSV
vatex_described.json	4,373	VATEX — captions synthesized via GPT-4o-mini

Video files are not included in this repo. Please download VATEX videos from the original source: lmms-lab/VATEX

Format

Each entry is a multi-turn chat message in the following structure:

{
  "messages": [
    {
      "role": "system",
      "content": "You are a video description assistant. Watch the video and answer with one clear natural sentence in lowercase describing the main visible action."
    },
    {
      "role": "user",
      "content": [
        {"type": "video", "video": "VATEX/<videoID>"},
        {"type": "text", "text": "Describe the main action happening in this video in one sentence."}
      ]
    },
    {
      "role": "assistant",
      "content": [{"type": "text", "text": "<one-sentence lowercase description>"}]
    }
  ],
  "label": "<action label>",
  "task_type": "video_description",
  "label_style": "natural_description",
  "source_format": "converted_from_vatex_captions"
}

Data Sources

Kinetics-40K (train_described.json)

• Videos from Kinetics-400
• Labels directly from k400_train/val/test.csv annotation files
• source_format: "converted_from_annotation_csv"

VATEX (vatex_described.json)

• Videos from lmms-lab/VATEX (Kinetics-600 subset)
• Overlap removal: 105 clips duplicated in Kinetics-40K were removed (4,478 → 4,373)
• Labels and descriptions synthesized from VATEX English captions via GPT-4o-mini
• 15 clips matched to k400 labels directly; remaining 4,358 use AI-generated labels
• source_format: "converted_from_vatex_captions"

Video Setup

Videos are not bundled in this repo. To prepare the full dataset:

1. Download VATEX videos from lmms-lab/VATEX
2. Place videos under VATEX/<videoID>.mp4 to match the "video" paths in vatex_described.json

Usage

Designed for SFT training with GemmaFT — Gemma 4 E4B video action recognition pipeline.

import json

with open("vatex_described.json") as f:
    dataset = json.load(f)

# Each entry has .messages (SFT format), .label, .task_type
print(dataset[0]["label"])

Model Target

• Model: google/gemma-4-e4b-it (Gemma4ForConditionalGeneration)
• Task: Video action recognition SFT
• Stage: Stage 1 LoRA fine-tuning (r=16, α=32)