saliency
Datasets
All datasets matching “saliency”video_saliency
SalTempto — Saliency Eye-Tracking Dataset
A video eye-tracking dataset of 224 naturalistic video stimuli with gaze recordings from multiple subjects, designed for training and evaluating visual saliency models.
Dataset Overview
Videos
224 (1920×1080, ~30 fps)
Train / Val / Test
204 / 10 / 10
Subjects per video
Train: 1–3 (mean ≈ 2.5) · Val: 15–16 · Test: held out
Gaze recordings for the test split are intentionally held out as a hidden… See the full description on the dataset page: https://huggingface.co/datasets/anonymous-saltempto-submission/video_saliency.qwen36-q4km-swebench-lite-saliency-overfit-layercompare
Qwen3.6 Q4_K_M SWE-bench Lite Saliency Layer Residency
This dataset contains the reduced-scope Q4_K_M-only comparison of:
attention10_baseline_nommap
hybrid_promote_5_b11_nommap
The run uses SWE-bench Lite oracle-context patch-generation prompts, Q4_K_M
MTP, 64k context, q8_0/q8_0 KV cache, Flash Attention, no-mmap, identical
prompt order, and identical sampling. Quality grading is deferred, but
prediction JSONLs are included for later official SWE-bench evaluation.
Read… See the full description on the dataset page: https://huggingface.co/datasets/jakeatx/qwen36-q4km-swebench-lite-saliency-overfit-layercompare.qwen36-moe-coding-expert-saliency
Qwen3.6 MoE Coding Expert Saliency Analysis
This repository contains the combined coding-task expert-saliency report for local Qwen3.6-35B-A3B MoE tracing work.
Primary artifact:
combined_coding_expert_analysis.html: combined expert activation, router-weight, contribution-proxy, correlation, and residency-candidate analysis across the SciCode MVP trace set and the full Terminal-Bench Hard real-agent trace run.
Companion artifacts:… See the full description on the dataset page: https://huggingface.co/datasets/jakeatx/qwen36-moe-coding-expert-saliency.GLM-5.2-MaxMin108-saliency
GLM-5.2 expert saliency and keep-lists
Everything needed to reproduce
turintech/GLM-5.2-MaxMin108-GGUF
from unsloth/GLM-5.2-GGUF: about fifteen
minutes of disk I/O, no GPU, no calibration run.
Contents
keep_lists/: which experts survive, per layer. keep_mix108_maxmin.json produces the
published model. The others are the ablations (code-only, general-only and random selection
at several K), four of which degenerate on out-of-domain prompts.
saliency_merged/:… See the full description on the dataset page: https://huggingface.co/datasets/turintech/GLM-5.2-MaxMin108-saliency.video_saliency_example
SalTempto — Saliency Eye-Tracking Dataset (Example Subset)
Note: This repository contains a small example subset of the full SalTempto video_saliency dataset (3 of 224 videos) for testing and development purposes. The 3 preview videos were drawn by uniform random sampling from the training and validation splits of the full dataset (2 from train, 1 from validation); the test split is not represented. The full dataset is available at anonymous-neurips-submission/video_saliency.… See the full description on the dataset page: https://huggingface.co/datasets/anonymous-saltempto-submission/video_saliency_example.perceptual-saliency-pairs