Video RLVR — final training data - a ngqtrung Collection

ngqtrung 's Collections

Code-Switching DPO (Interspeech 2026)

Video RLVR — final training data

Qwen3-VL-8B video RLVR - GRPO checkpoint ladder

Qwen3-VL-8B RLVR — Models (v1)

Qwen3-VL-8B RLVR — Datasets (v1)

VMAR — Raw & Source

VMAR — Train-Ready (SFT + RL + Eval)

Video RLVR — final training data

updated about 1 month ago

The datasets behind our Qwen3-VL-8B video RLVR runs: the base 24f/100k mixture and the HopChain v5 multi-hop corpora.

ngqtrung/hopchain-v5

Viewer • Updated Sep 2 • 23.6k • 103

Note HopChain v5 — CURRENT. 22,550 train / 1,000 curated val, one question per video.
ngqtrung/video-r1-mc-24f100k-combined

Viewer • Updated Sep 2 • 106k • 130

Note Video-R1 MC 24f/100k — 105,993 prompts. The base GRPO mixture the v5 runs start from.