Commit History
-
Remove the hero subtitle
Jonna Marie Matthiesen +1: Claude Fable 5Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AGJFdqmf4AXcsYvb56LLCC336b0fe -
Default each family's chart to its best-improvement metric
Jonna Marie Matthiesen +1: Claude Fable 5A configured chart default_metric (family or global) still wins, and an explicitly picked metric (button click or #metric= link) is pinned across family switches. Strip the stale ips defaults from the CV families so the automatic pick applies there. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AGJFdqmf4AXcsYvb56LLCCc665bdc -
Place VS BASE column directly after the metric it compares
Jonna Marie Matthiesen +1: Claude Fable 5Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01AGJFdqmf4AXcsYvb56LLCC9962162
-
Add shareable URL state, speedup column and family-scoped metric legend (#7)
- Add shareable URL state, speedup column and family-scoped metric legend (dbb911ba256b7aa68e261eb5899d1a86374733c9)8d85328
-
RF-DETR demo: lead with pedestrian street detection video
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>e9e9bbc -
Embed video demos for RF-DETR Base and DINOv3
Threshold-sweep and patch-similarity anchor-sweep MP4s, both rendered from real FP16/INT8 TensorRT engines on Jetson devices. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>3d4f9cc -
Add SAM-3D-Body mesh-fidelity table from the model card
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>b1cd606 -
Add accuracy tables for five families missing them
From each model card: all-MiniLM-L6-v2 (STS-B), MobileViT-Small (ImageNet), paraphrase-multilingual-MiniLM-L12-v2 (multilingual STS), SAM3 (SA-Co/Gold), Chronos-2 (WQL ratio, lower-is-better flag). Also point RF-DETR accuracy_url at the renamed repo. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>93c3d46
-
Fix remaining metric header cell
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>fdab5db -
Readable metric headers for RF-DETR Base accuracy table
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>1b0c494 -
Add RF-DETR Base to Edge Inference Benchmarks
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>745dcd1
-
SAM-3D-Body: disclose encoder-only scope in setup note
89e6098 -
SAM3: correct AGX Orin rows to model-card table values
a20439e -
Parakeet: correct AGX numbers, add L4 device + peak-memory metric
6657987
-
Support ?family= deep links (model cards link to their own family)
133b3c7 -
Add deploy-benchmark model families (8 models, lat/ips/mem metrics)
45c0521
-
Add FlashHead version to measurement setup
Jonna Marie Matthiesen9560ed1 -
Add NVIDIA prefix to Device names
Jonna Marie Matthiesen1550e5f
-
Qwen3.5-FlashHead (#6)
- Remove BS=8 benchmarks (e3da2051b465507f3c26f673d44f9a2501585045) - Add Qwen3.5 accuracy data (a8408e4835493ffe2ddc5c809d80b1fbcced8021) - Add Qwen3.5 latency data (59808f20e8d2005fd81bf75bc2ff1bf2dccbe271)d4554af
-
Move OOM rows to bottom
Jonna Marie Matthiesen5df6c82 -
Fix bar plot OOM for Cosmos Reason2
Jonna Marie Matthiesendb3c303 -
Update footer
Jonna Marie Matthiesene0c5f69
-
Show N/A for unmeasured TPOT/TTFT and hide all-N/A columns
Jonna Marie Matthiesen +1: Claude Opus 4.6- Replace ttft=0.00 and tpot with N/A in Llama-3.2, Gemma-3, and Qwen3 text benchmarks where these metrics were not actually measured - Update CSV parser to preserve N/A as a string - Hide metric columns from table and chart when all values are N/A Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>c73bd71 -
Add accuracy tables, RTX 3500 Ada latency, and fix 270M variant grouping
Jonna Marie Matthiesen +1: Claude Opus 4.6- Add accuracy CSVs for Llama-3.2, Gemma-3, and Qwen3 (baseline vs FlashHead) - Add RTX 3500 Ada TPS-only latency data for all families with README measurements - Add Gemma-3-270M RTX 3500 Ada data (first latency data for this variant) - Register rtx_3500_ada device in config with experiment setup - Fix deriveBaseFamily regex to match M (million) suffixes so Gemma-3-270M groups under Gemma-3 in the sidebar Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>7671b7c -
Group table rows by model sub-family (size) instead of individual model
Jonna Marie Matthiesen +1: Claude Opus 4.6 (1M context)Changes table_group_by from "model" to "model_family" so that general families like Llama-3.2 visually group by size (1B, 3B) with separators and per-group best-metric highlighting. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>bd14011 -
Extract embed SVG generator into embed.js and add version tracking
Jonna Marie Matthiesen +1: Claude Opus 4.6 (1M context)Move the embed-in-model-card feature (SVG generation, modal, upload workflow) from app.js into a standalone embed.js module. Remove HF token auto-upload logic, keeping only the manual Download & Open HF workflow. Add data-commit metadata to generated SVGs for staleness detection. Use variant-aware filenames so Llama-3.2-1B and 3B get distinct SVGs. Right-align metric buttons in chart header. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>5969585 -
Add Llama-3.2, Gemma-3, and Qwen3 benchmarks and improve chart rendering
Jonna Marie Matthiesen +1: Claude Opus 4.6 (1M context)Add benchmark data for three new model families (Llama-3.2, Gemma-3, Qwen3) on NVIDIA edge devices using vLLM 0.19.0. Improve chart/table UX by sorting bars by model size, hiding all-zero metrics, prefixing labels when multiple families are shown, and supporting configurable base display names. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>b96722d
-
Fix Nano video OOM rows: one row per model, match image format
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>7faf78c -
Remove Orin Nano 1920x1080 12-frame OOM rows
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>f1d787a -
Add nvidia/Cosmos-Reason2-2B baseline 12-frame video benchmarks
AGX Orin and AGX Thor, 854x480 / 1280x720 / 1920x1080, batch=1, processor-fps=4 for actual 12-frame inference. Nano already marked OOM. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>44a2d55 -
Add Nano OOM entries for nvidia/Cosmos-Reason2-2B 12-frame video
Baseline model is too large for Orin Nano Super at gpu_util=0.75. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>381a991 -
Replace 12-frame video benchmarks with correct processor-fps=4 results
Old 12-frame entries used the processor's default fps=2, making them effectively 6-frame benchmarks mislabeled as 12. New results use --processor-fps 4 for actual 12-frame inference. Benchmarked: W4A16, Edge2, FlashHead on AGX Thor, AGX Orin, Orin Nano Super at 854x480, 1280x720, 1920x1080. Orin Nano 1920x1080 = OOM. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>c8b2bc9
-
Remove NVFP4A16 and W4A16 from Physical AI Bench
d26b351 -
Update Physical AI Bench for Cosmos-Reason2-2B and add FlashHead
100a273
-
Update image data and add FlashHead benchmarks
Jonna Marie Matthiesen9d40ec2 -
Fix video loading time, prebuffer vids
Jonna Marie Matthiesen234ef1f -
Demo (#5)
- Add demo (34e8f9e98231f0669c21fb1a0a7d2eb9cdad8dce)4db51d5
-
Add accuracy data (#4)
- Add accuracy table above filters for model families with accuracy data (19a0613a6092b8c537777932e8838513d444d366) - Make model column sticky when scrolling tables horizontally (05b2588a0614da2b43266f38c01e954e03f84db7) - Remove background from sticky model column (8b6d8a777e700cfe91758da1c7662e12d5b8af15) - Set sticky model column background to page base color (79e2f2a1088b06ec8d89331d2a741a943bd57cce) - Split accuracy table into fixed model column and scrollable metrics (19914a87f67066fba34de73af85b251767495f5f) - Add spacing below accuracy table (160dcdd285c41f15daec0d0d69cd1badcdb75b83) - Add configurable accuracy table with title, link, and data (949674753f50eb26e4591812269899f07bc9790e)b36fa96
-
Add Cosmos-Reason2-2B-W4A16-Edge2-FlashHead benchmark data
Verified signature28530a0
-
Add Qwen3.5 Orin Nano benchmarks
Jonna Marie Matthiesen2d4b81f -
Add subgroups to sidebar (#3)
- Add hierarchical sidebar grouping models by base family (4d94a6cb8b86643f957a95672b4a63ba2f4674fe) - Refresh device filter buttons when switching variants (74ad70e95ade723232cf0da29fc6180044c219f6) - Add default_device config option per model family (ecc183ea173d0c6ff301ddf338ec7806b305404c) - Apply default_device when switching between base families (9527fc8ccc7e23b203a45c3ac1cef7cedce4debb) - Use descriptive names when multiple models would share the same chart label (5e234e6186a43a01e0ed2ef3c3423c2397cb87b9)d285c40
-
Change neutral from #667788 to #98A6AF
Verified signature6156d55 -
Add Qwen3.5 Benchmarks (#2)
- Make filter names all CAPS (d0e8c47f09dde7b3437a7463e126ae7a6ee80059) - Order tables by device order (3ba0d66d5e57da46fc61a354ab671e477a8212ce) - Add Qwen3.5 AGX Thor benchmarks (e36fba8de41d0ba90424927064fa7b02b61e3dcc) - Add Qwen3.5 Orin benchmarks (f60dbfde85e4bd0039a0990560c80cf9100e3ece)feca03d -
Update chart and table styling (#1)
- Update chart and table styling (5d8363578c54b70d46817f8940be1b1f3209fd70)45f0a42
-
Make radial-gradient more visible
Verified signature2103ab2 -
Update app.js
Verified signature60880be -
Update page
Verified signature32853d9 -
Update data/cosmos-reason2.csv
Verified signature979f9e7 -
Rename data/cosmos_reason2.csv to data/cosmos-reason2.csv
Verified signature97bb278 -
Delete data/cosmos-reason2.csv
Verified signature969b2b3