-> **MedChain-R1-4B** is the 4B sibling of MedChain-R1-9B, a vision-language model for sequential medical image grounding built on Qwen3.5-4B ("Think Before You Locate", arXiv:2611.00041). Given a sequence of medical images and a clinical question, it first writes an image-grounded rationale in `<think>` tags, then returns one pixel-coordinate bounding box. Training used deliberative chain-of-thought supervision (dCoT-SFT) on the MedChain corpus, then GRPO with an IoU plus format reward. MedChain has 182,573 filtered rationales over the 8 MedSG-Bench tasks and 10 imaging modalities. A medical VLM (FlemingVL-38B) drafts each rationale and a multimodal refiner (Qwen3.5-397B-A17B) rewrites it conditioned on the annotated box. This checkpoint (dCoT-SFT checkpoint-2000 followed by one GRPO pass, seed 42) scores 75.97 mIoU / 83.67 Acc@0.5 on all 9,630 MedSG-Bench test questions. That is slightly above the paper's 4B GRPO row (75.32 / 82.89, from a different checkpoint) and above the supervised state of the art MedSeq-Grounder 7B (72.55 / 79.71). It is below the 9B sibling (77.54 / 84.87), which the card recommends for highest accuracy, while the 4B is pitched at resource-constrained settings. The paper notes the gains vary by task. Against an answer-only 4B model, the reasoning supervision gains 14.7–16.1 mIoU on difference grounding but loses 2.2–3.2 on multi-view, patch and referring grounding, and GRPO adds little after dCoT-SFT. The card says to append its exact instruction to every question, generate one example at a time because batching gave degenerate outputs, and use the bundled processor, since a lower-resolution one costs 3–5 points. The rationales were written knowing the answer, so they may be post-hoc rather than a faithful account of the model's reasoning. It is a research artifact, not validated for diagnosis or clinical use, and is evaluated only on MedSG-Bench. The license is Apache 2.0, and the training data is released as MedChain-R1-data.
30
31
> [!NOTE]
32
> MTP — Speculative Decoding: No
26
27
# **MedChain-R1-4B-GGUF**
28
29
+> **MedChain-R1-4B** is the 4B sibling of MedChain-R1-9B, a vision-language model for sequential medical image grounding built on Qwen3.5-4B ("Think Before You Locate"). Given a sequence of medical images and a clinical question, it first writes an image-grounded rationale in `<think>` tags, then returns one pixel-coordinate bounding box. Training used deliberative chain-of-thought supervision (dCoT-SFT) on the MedChain corpus, then GRPO with an IoU plus format reward. MedChain has 182,573 filtered rationales over the 8 MedSG-Bench tasks and 10 imaging modalities. A medical VLM (FlemingVL-38B) drafts each rationale and a multimodal refiner (Qwen3.5-397B-A17B) rewrites it conditioned on the annotated box. This checkpoint (dCoT-SFT checkpoint-2000 followed by one GRPO pass, seed 42) scores 75.97 mIoU / 83.67 Acc@0.5 on all 9,630 MedSG-Bench test questions. That is slightly above the paper's 4B GRPO row (75.32 / 82.89, from a different checkpoint) and above the supervised state of the art MedSeq-Grounder 7B (72.55 / 79.71). It is below the 9B sibling (77.54 / 84.87), which the card recommends for highest accuracy, while the 4B is pitched at resource-constrained settings. The paper notes the gains vary by task. Against an answer-only 4B model, the reasoning supervision gains 14.7–16.1 mIoU on difference grounding but loses 2.2–3.2 on multi-view, patch and referring grounding, and GRPO adds little after dCoT-SFT. The card says to append its exact instruction to every question, generate one example at a time because batching gave degenerate outputs, and use the bundled processor, since a lower-resolution one costs 3–5 points. The rationales were written knowing the answer, so they may be post-hoc rather than a faithful account of the model's reasoning. It is a research artifact, not validated for diagnosis or clinical use, and is evaluated only on MedSG-Bench. The license is Apache 2.0, and the training data is released as MedChain-R1-data.
