We just released TRL v0.20 with major multimodal upgrades!
👁️ VLM support for GRPO (highly requested by the community!) 🎞️ New GSPO trainer (from @Qwen, released last week, VLM-ready) 🐙 New MPO trainer (multimodal by design, as in the paper)
🧑🍳 In this @HuggingFace Face Cookbook notebook, we demonstrate how to align a multimodal model (VLM) using Mixed Preference Optimization (MPO) using trl.
💡 This recipe is powered by the new MPO support in trl, enabled through a recent upgrade to the DPO trainer!
We align the multimodal model using multiple optimization objectives (losses), guided by a preference dataset (chosen vs. rejected multimodal pairs).
Many VLMs claim to process hours of video. But can they follow the story?🤔 Today, we introduce TimeScope: The benchmark that separates true temporal understanding from marketing hype. Let's see how much VLMs really understand!⏳
We test three skills that matter for real-world use: 🔎 Localized Retrieval: Find a specific action. 🧩 Information Synthesis: Piece together scattered clues. 🏃 Fine-Grained Perception: Analyze detailed motion (e.g., count how many times a person swings an axe).
The results are in, and they're revealing. Only Gemini 2.5 pro handles 1-hour-long videos. Performance drops sharply with duration, proving that long video understanding is still challenging. We've found the breaking points—now the community can start fixing them.📈
Want to learn more? TimeScope is 100% open-source. Benchmark your model and help us build the next generation of video AI.
