Video-HopChain-8B-GGUF

Video-HopChain-8B by ngqtrung (Video-HopChain: Multi-Hop Questions and Confidence-Gated Exploration for Video Reasoning Models) is a Qwen3-VL-8B-Instruct checkpoint trained with GRPO in two stages — first on a 105,993-row general video QA mixture, then on the Video-HopChain dataset with Confidence-Gated Exploration (CGE) — representing the full "standard RL + Video-HopChain + CGE" configuration from Table 1 of the paper. CGE addresses advantage collapse in GRPO: when the first 4 of 8 rollouts per question are all correct or all incorrect (yielding zero reward variance), the remaining 4 rollouts are drawn with the policy's top token masked wherever its probability exceeds 0.95 within the reasoning span, forcing exploration without consuming extra rollout budget. Initialized from the standard-RL stage and further trained on 22,550 Video-HopChain rows at 140 frames, the model reasons inside <think> tags and outputs a boxed final answer, achieving the best results in the paper's comparison across eight video benchmarks — a 59.3% mean accuracy versus 52.3% for the base Qwen3-VL-8B-Instruct and 57.9% for the standard-RL-plus-Video-HopChain variant without CGE, with particularly strong gains on the in-domain held-out split (23.2% vs 13.4% for the base model) and on VCRBench (44.9%) and PerceptionComp (37.4%). It's loadable via standard Transformers (Qwen3VLForConditionalGeneration) in BF16, was trained on 4 nodes of 8 H100 80GB GPUs using an asynchronous verl trainer, and is released under the Apache 2.0 license; the authors note it was evaluated at a single model scale and that the training dataset is synthetic and caption-verified rather than frame-verified, so caption errors can propagate into labels.

Citation
Video-HopChain-8B

Model Files

File Name	Quant Type	File Size	File Link	Description
Video-HopChain-8B.BF16.gguf	BF16	16.4 GB	Link	Full BF16 weights. Highest quality, largest file size.
Video-HopChain-8B.Q3_K_L.gguf	Q3_K_L	4.43 GB	Link	Lower quality but usable, good for low RAM availability.
Video-HopChain-8B.Q3_K_M.gguf	Q3_K_M	4.12 GB	Link	Low quality.
Video-HopChain-8B.Q4_K_M.gguf	Q4_K_M	5.03 GB	Link	Good quality, default size for most use cases, recommended.
Video-HopChain-8B.Q4_K_S.gguf	Q4_K_S	4.8 GB	Link	Slightly lower quality with more space savings, recommended.
Video-HopChain-8B.Q5_K_M.gguf	Q5_K_M	5.85 GB	Link	High quality, recommended.
Video-HopChain-8B.Q5_K_S.gguf	Q5_K_S	5.72 GB	Link	High quality, recommended.
Video-HopChain-8B.Q6_K.gguf	Q6_K	6.73 GB	Link	Very high quality, near perfect, recommended.
Video-HopChain-8B.mmproj-bf16.gguf	mmproj-bf16	1.16 GB	Link	Multimodal projection file in BF16 format. Used for vision/language models.

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp