SAGE-MM-Qwen2.5-VL-7B-SFT_RL-GGUF

The SAGE-MM-Qwen2.5-VL-7B-SFT_RL from allenai is a 7B-parameter reinforcement learning (RL) refined vision-language model, further post-trained via RL from the SAGE-MM-Qwen2.5-VL-7B-SFT (itself fine-tuned from Qwen/Qwen2.5-VL-7B-Instruct), acting as the core decision-maker in the SAGE (Smart Any-Horizon Agent) system for long video reasoning with enhanced performance over SFT baselines through two-stage operation: Stage-1 assesses initial sampled frames/metadata to route queries as single-turn (direct answers) or multi-turn (tool-needed), while Stage-2 loops JSON-formatted tool calls for web-search, timestamped ASR transcription, event grounding, video frame/subclip extraction, and visual analysis to iteratively build context until resolution. Tailored for arbitrary-length video Q&A across sports, narratives, events, or timelines beyond fixed horizons, it demands the SAGE GitHub runtime for tool parsing/execution/observation feedback, delivering state-of-the-art results on benchmarks like MINERVA under Apache 2.0 for research/educational use per Ai2 guidelines, with GGUF quantizations enabling efficient deployment.

SAGE-MM-Qwen2.5-VL-7B-SFT_RL [GGUF]

Quants Usage