Domain SFT teachers, GSPO teachers and the training data (SFT v4, RL v2 with full media) of the MOPD project.
Nguyen Quang Trung
ngqtrung
·
AI & ML interests
None yet
Recent Activity
updated a collection 4 days ago
MOPD: multi-teacher on-policy distillation (Qwen3.5-9B) updated a collection 4 days ago
MOPD: multi-teacher on-policy distillation (Qwen3.5-9B) updated a collection 4 days ago
MOPD: multi-teacher on-policy distillation (Qwen3.5-9B)Organizations
Multi-hop video questions for RLVR, and the 8B model trained on them with Confidence-Gated Exploration.
-
ngqtrung/Video-HopChain
Viewer • Updated • 23.6k • 425 -
ngqtrung/Video-HopChain-8B
Video-Text-to-Text • 9B • Updated • 47 • 1 -
ngqtrung/Video-HopChain-8B-Standard-RL
Video-Text-to-Text • 9B • Updated • 30 -
Video-HopChain: Multi-Hop Questions and Confidence-Gated Exploration for Video Reasoning Models
Paper • 2609.25773 • Published • 1
Every published checkpoint of the 8B video-QA GRPO campaign, ranked by core-3 mean_accuracy (5,645 rows). Untrained base = 0.4426.
Curated SFT + GRPO RL datasets (video MC-QA, OMR math-image, OpenMMReasoner-RL, Vero) for Qwen3-VL-8B post-training.
Train-ready VMAR datasets: teacher-distilled SFT corpus, RL prompt set, and the curated in-loop eval benchmark.
Interspeech 2026 paper + DPO data for EN-ZH code-switching ASR. Page: ngquangtrung57.github.io/codeswitch-dpo-page
The datasets behind our Qwen3-VL-8B video RLVR runs: the base 24f/100k mixture and the HopChain v5 multi-hop corpora.
Qwen3-VL-8B GRPO RLVR checkpoints from a token-dropout exploration study. OMR ppexplore=winner (0.714); video ~0.485 dead-heat.
Raw multi-style distilled traces, the pre-distillation template seed, and the never-trained real-audio eval.
Domain SFT teachers, GSPO teachers and the training data (SFT v4, RL v2 with full media) of the MOPD project.
Interspeech 2026 paper + DPO data for EN-ZH code-switching ASR. Page: ngquangtrung57.github.io/codeswitch-dpo-page
Multi-hop video questions for RLVR, and the 8B model trained on them with Confidence-Gated Exploration.
-
ngqtrung/Video-HopChain
Viewer • Updated • 23.6k • 425 -
ngqtrung/Video-HopChain-8B
Video-Text-to-Text • 9B • Updated • 47 • 1 -
ngqtrung/Video-HopChain-8B-Standard-RL
Video-Text-to-Text • 9B • Updated • 30 -
Video-HopChain: Multi-Hop Questions and Confidence-Gated Exploration for Video Reasoning Models
Paper • 2609.25773 • Published • 1
The datasets behind our Qwen3-VL-8B video RLVR runs: the base 24f/100k mixture and the HopChain v5 multi-hop corpora.
Every published checkpoint of the 8B video-QA GRPO campaign, ranked by core-3 mean_accuracy (5,645 rows). Untrained base = 0.4426.
Qwen3-VL-8B GRPO RLVR checkpoints from a token-dropout exploration study. OMR ppexplore=winner (0.714); video ~0.485 dead-heat.
Curated SFT + GRPO RL datasets (video MC-QA, OMR math-image, OpenMMReasoner-RL, Vero) for Qwen3-VL-8B post-training.
Raw multi-style distilled traces, the pre-distillation template seed, and the never-trained real-audio eval.
Train-ready VMAR datasets: teacher-distilled SFT corpus, RL prompt set, and the curated in-loop eval benchmark.