AnvitaPrasad/vgate-outputs

VGate — per-query model outputs

Raw outputs behind the results of VGate: Auditing What Retrieval Gates Learn for Vision-Language Models (code: https://github.com/AnvitaPrasad/vgate; hosted log view: https://huggingface.co/spaces/AnvitaPrasad/vgate).

One folder per GPU job (J1 … J23). Each holds, per benchmark and reader:

• <benchmark>__<reader>__<condition>.jsonl: one line per question with the question id, the model's reply, token log-probabilities and entropies, prompt and reply lengths and GPU seconds. Conditions: ne closed-book, ev with the retrieved evidence, oracle with the gold evidence, and auxiliary passes (samples, published gate prompts, ...).
• hidden_<benchmark>__<reader>__L<layer>_part<k>.npy with …__qids<k>.json: the last prompt token's hidden state of the closed-book pass at three layers (float16).
• meta__*.json: model, precision, prompt id and decoding settings of the run.

Readers: Qwen3-VL-4B-Instruct, Qwen2.5-VL-3B-Instruct, InternVL3.5-4B, Qwen3-VL-8B-Instruct, and in J23 Qwen3-VL-32B-Instruct with 4-bit weights on 500 held-out questions per benchmark; J17 holds the retrieval-token decisions of the released ReflectiVA checkpoint. In J18, 72 questions were generated in float32 after failing in float16 (see the repository's REPRODUCIBILITY.md).

No benchmark content is included: no question text, images, gold answers or retrieved passages, only question ids and model outputs. Rebuild the question tables from the benchmarks' own sources as described in the repository.

Use: python scripts/fetch_outputs.py in the repository places these folders under experiments/runs/, after which make score confirm extend external cvqa cvqa2 results regenerates every result file.

Licence

These files are model outputs (replies, log-probabilities, hidden states). They follow the terms of the model that produced them: Qwen3-VL-4B / 8B / 32B-Instruct, Qwen3-VL-4B-Thinking and InternVL3.5-4B outputs, Apache-2.0; Qwen2.5-VL-3B-Instruct outputs (the files named *Qwen2.5-VL-3B* in J2, J6, J18, J19, J20), Qwen Research License: non-commercial use only; ReflectiVA's decisions (J17), Apache-2.0. The MIT licence of the code repository does not cover them. The dataset viewer is off because the files are raw per-run outputs with different columns, not one table.