Add Qwen3.8-27B-FP8 twostep_single full benchmark_312 result (14,896 requests)

.gitattributes

@@ -58,3 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+benchmark_312/qwen3.8-27b-fp8_twostep_single/responses.jsonl filter=lfs diff=lfs merge=lfs -text

benchmark_312/qwen3.8-27b-fp8_twostep_single/README.md

@@ -0,0 +1,69 @@
+# Qwen3.8-27B-FP8 · twostep_single prompt · benchmark_312 (full, 14,896 requests)
+
+Zero-shot point correspondence on the frozen 312-clip release (`2inf/tracking_bench_vlm`,
+`releases/benchmark_312_chen_v1`): 312 clips, 14,896 requests, each one reference image with one marked point and one
+unmarked target frame. One model call per request, no thinking mode, temperature 0, no post-processing.
+
+Run folder in the lab repository: `experiments/benchmark_312_twostep_qwen38/v1`, finished 2026-10-01.
+
+## Score (TAP-Vid metrics, offsets scaled to 256 x 256)
+
+| split | queries | visible | <1 px | <2 px | <4 px | <8 px | <16 px | d_avg | AJ | OA |
+|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| all | 14896 | 11662 | 2.45 | 7.19 | 17.74 | 35.14 | 58.27 | **24.16** | 13.52 | 78.31 |
+| drivetrack | 1221 | 897 | 1.56 | 6.47 | 21.40 | 49.16 | 73.69 | **30.46** | 17.66 | 72.07 |
+| drivtap | 486 | 348 | 4.89 | 10.92 | 37.93 | 65.23 | 82.47 | **40.29** | 24.45 | 78.40 |
+| egopoints | 857 | 857 | 4.20 | 10.50 | 20.07 | 34.77 | 55.78 | **25.06** | 15.81 | 99.07 |
+| itto | 1200 | 885 | 2.82 | 8.47 | 18.98 | 34.58 | 54.80 | **23.93** | 12.42 | 73.58 |
+| perception_test | 1206 | 1206 | 4.31 | 12.27 | 25.70 | 46.27 | 69.24 | **31.56** | 21.75 | 97.01 |
+| robotap | 1200 | 854 | 1.17 | 4.45 | 12.30 | 28.69 | 50.94 | **19.51** | 10.09 | 71.33 |
+| robotrack_real | 1022 | 770 | 5.32 | 13.25 | 25.97 | 45.58 | 70.52 | **32.13** | 18.40 | 76.32 |
+| tapvid3d_adt | 1207 | 962 | 0.52 | 2.91 | 11.95 | 28.48 | 52.39 | **19.25** | 11.01 | 79.87 |
+| tapvid3d_pstudio | 1200 | 930 | 0.97 | 3.33 | 10.86 | 28.28 | 58.06 | **20.30** | 11.31 | 75.58 |
+| tapvid_davis | 1102 | 813 | 2.71 | 8.73 | 25.71 | 48.34 | 72.57 | **31.61** | 17.94 | 75.95 |
+| tapvid_kinetics | 1200 | 861 | 1.86 | 7.67 | 20.21 | 42.04 | 66.78 | **27.71** | 15.41 | 74.25 |
+| tapvid_kubric | 1200 | 932 | 0.21 | 0.64 | 2.25 | 8.48 | 30.04 | **8.33** | 4.18 | 76.75 |
+| tapvid_rgb_stacking | 1200 | 864 | 0.12 | 1.39 | 3.94 | 11.92 | 34.03 | **10.28** | 5.02 | 71.42 |
+| vl_surgpt | 160 | 120 | 3.33 | 7.50 | 21.67 | 39.17 | 56.67 | **25.67** | 14.33 | 80.00 |
+| votsp2026 | 435 | 363 | 8.82 | 18.46 | 30.30 | 41.60 | 60.88 | **32.01** | 18.25 | 82.99 |
+
+By frame gap between the reference and the target frame:
+
+| split | queries | visible | <1 px | <2 px | <4 px | <8 px | <16 px | d_avg | AJ | OA |
+|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
+| gap 1 | 360 | 266 | 8.27 | 20.30 | 44.36 | 72.56 | 87.59 | **46.62** | 28.43 | 72.78 |
+| gap 2-15 | 4798 | 3698 | 2.57 | 7.84 | 19.93 | 40.45 | 65.44 | **27.25** | 15.46 | 77.64 |
+| gap 16-30 | 2567 | 1964 | 1.88 | 6.47 | 16.24 | 33.30 | 56.47 | **22.87** | 12.50 | 76.00 |
+| gap 31-60 | 2659 | 2034 | 2.11 | 6.24 | 16.13 | 31.91 | 55.90 | **22.46** | 12.34 | 76.57 |
+| gap 61-120 | 2120 | 1608 | 2.11 | 5.60 | 15.05 | 29.10 | 50.56 | **20.49** | 11.01 | 76.13 |
+| gap 121+ | 2392 | 2092 | 2.63 | 7.22 | 15.54 | 30.50 | 51.82 | **21.54** | 12.58 | 86.83 |
+
+- d_k: share of the visible targets whose answer lies within k px at the 256 x 256 scale; d_avg = mean over k in 1, 2, 4, 8, 16.
+- AJ: average Jaccard over the same thresholds (a far answer on a visible point counts as a false positive and a false negative; a position given for an occluded point is a false positive).
+- OA: occlusion accuracy over all queries.
+- Answered 14878 of 14896 requests (18 unparsable replies, scored as wrong).
+- Tokens: 35,797,069 input, 1,394,783 output, 0 reasoning.
+
+## Setting
+
+| item | value |
+|---|---|
+| model | Qwen/Qwen3.8-27B-FP8 served with vLLM (bfloat16, `enable_thinking: false`, `max_tokens` 2000, temperature 0) |
+| image input | native resolution, `min_pixels` 65,536, `max_pixels` 4,194,304 |
+| coordinates | the model answers on the 0-1000 scale of each image; the parser converts to pixels (`norm1000` convention) |
+| prompt | `prompt_twostep_single.txt` (two-step: first the crosshair position in IMAGE 1, then the point in the target frame) |
+| harness | `vlm_track.harness:single_call`, one call per request |
+| evaluator | `vlm_track.evaluation.tapvid:score` |
+
+## Files
+
+| file | content |
+|---|---|
+| `metrics.json` | the scores above, with the per-dataset and per-gap breakdowns and token usage |
+| `predictions.jsonl` | one line per request: `id`, `x`, `y` in native pixels of the target frame, `occluded` |
+| `responses.jsonl` | the raw model reply of every request (`calls[].text`) with timing and token usage |
+| `resolved.yaml` | the complete experiment configuration as run (model, serving, data, prompt text, harness) |
+| `prompt_twostep_single.txt` | the prompt template |
+
+Request ids are `<dataset>/<clip>/p<track>/r<reference frame>/f<target frame>` and match `requests.jsonl` /
+`truth.jsonl` of the release.

benchmark_312/qwen3.8-27b-fp8_twostep_single/metrics.json

@@ -0,0 +1 @@
+{"samples": 14896, "evaluation_items": 14896, "answered": 14878, "failed_samples": 18, "missing_predictions": 18, "calls": 14896, "response_records": 14896, "superseded_records": 0, "wall_seconds": null, "calls_with_usage": 14896, "usage": {"input_tokens": 35797069, "output_tokens": 1394783, "reasoning_tokens": 0}, "scores": {"queries": 14896, "answered": 14878, "visible_queries": 11662, "d_1": 2.45, "d_2": 7.19, "d_4": 17.74, "d_8": 35.14, "d_16": 58.27, "d_avg": 24.16, "aj": 13.52, "oa": 78.31}, "by_dataset": {"drivetrack": {"queries": 1221, "answered": 1206, "visible_queries": 897, "d_1": 1.56, "d_2": 6.47, "d_4": 21.4, "d_8": 49.16, "d_16": 73.69, "d_avg": 30.46, "aj": 17.66, "oa": 72.07}, "drivtap": {"queries": 486, "answered": 486, "visible_queries": 348, "d_1": 4.89, "d_2": 10.92, "d_4": 37.93, "d_8": 65.23, "d_16": 82.47, "d_avg": 40.29, "aj": 24.45, "oa": 78.4}, "egopoints": {"queries": 857, "answered": 857, "visible_queries": 857, "d_1": 4.2, "d_2": 10.5, "d_4": 20.07, "d_8": 34.77, "d_16": 55.78, "d_avg": 25.06, "aj": 15.81, "oa": 99.07}, "itto": {"queries": 1200, "answered": 1199, "visible_queries": 885, "d_1": 2.82, "d_2": 8.47, "d_4": 18.98, "d_8": 34.58, "d_16": 54.8, "d_avg": 23.93, "aj": 12.42, "oa": 73.58}, "perception_test": {"queries": 1206, "answered": 1206, "visible_queries": 1206, "d_1": 4.31, "d_2": 12.27, "d_4": 25.7, "d_8": 46.27, "d_16": 69.24, "d_avg": 31.56, "aj": 21.75, "oa": 97.01}, "robotap": {"queries": 1200, "answered": 1200, "visible_queries": 854, "d_1": 1.17, "d_2": 4.45, "d_4": 12.3, "d_8": 28.69, "d_16": 50.94, "d_avg": 19.51, "aj": 10.09, "oa": 71.33}, "robotrack_real": {"queries": 1022, "answered": 1021, "visible_queries": 770, "d_1": 5.32, "d_2": 13.25, "d_4": 25.97, "d_8": 45.58, "d_16": 70.52, "d_avg": 32.13, "aj": 18.4, "oa": 76.32}, "tapvid3d_adt": {"queries": 1207, "answered": 1207, "visible_queries": 962, "d_1": 0.52, "d_2": 2.91, "d_4": 11.95, "d_8": 28.48, "d_16": 52.39, "d_avg": 19.25, "aj": 11.01, "oa": 79.87}, "tapvid3d_pstudio": {"queries": 1200, "answered": 1200, "visible_queries": 930, "d_1": 0.97, "d_2": 3.33, "d_4": 10.86, "d_8": 28.28, "d_16": 58.06, "d_avg": 20.3, "aj": 11.31, "oa": 75.58}, "tapvid_davis": {"queries": 1102, "answered": 1102, "visible_queries": 813, "d_1": 2.71, "d_2": 8.73, "d_4": 25.71, "d_8": 48.34, "d_16": 72.57, "d_avg": 31.61, "aj": 17.94, "oa": 75.95}, "tapvid_kinetics": {"queries": 1200, "answered": 1200, "visible_queries": 861, "d_1": 1.86, "d_2": 7.67, "d_4": 20.21, "d_8": 42.04, "d_16": 66.78, "d_avg": 27.71, "aj": 15.41, "oa": 74.25}, "tapvid_kubric": {"queries": 1200, "answered": 1200, "visible_queries": 932, "d_1": 0.21, "d_2": 0.64, "d_4": 2.25, "d_8": 8.48, "d_16": 30.04, "d_avg": 8.33, "aj": 4.18, "oa": 76.75}, "tapvid_rgb_stacking": {"queries": 1200, "answered": 1200, "visible_queries": 864, "d_1": 0.12, "d_2": 1.39, "d_4": 3.94, "d_8": 11.92, "d_16": 34.03, "d_avg": 10.28, "aj": 5.02, "oa": 71.42}, "vl_surgpt": {"queries": 160, "answered": 159, "visible_queries": 120, "d_1": 3.33, "d_2": 7.5, "d_4": 21.67, "d_8": 39.17, "d_16": 56.67, "d_avg": 25.67, "aj": 14.33, "oa": 80.0}, "votsp2026": {"queries": 435, "answered": 435, "visible_queries": 363, "d_1": 8.82, "d_2": 18.46, "d_4": 30.3, "d_8": 41.6, "d_16": 60.88, "d_avg": 32.01, "aj": 18.25, "oa": 82.99}}, "by_gap": {"1": {"queries": 360, "answered": 360, "visible_queries": 266, "d_1": 8.27, "d_2": 20.3, "d_4": 44.36, "d_8": 72.56, "d_16": 87.59, "d_avg": 46.62, "aj": 28.43, "oa": 72.78}, "2-15": {"queries": 4798, "answered": 4793, "visible_queries": 3698, "d_1": 2.57, "d_2": 7.84, "d_4": 19.93, "d_8": 40.45, "d_16": 65.44, "d_avg": 27.25, "aj": 15.46, "oa": 77.64}, "16-30": {"queries": 2567, "answered": 2566, "visible_queries": 1964, "d_1": 1.88, "d_2": 6.47, "d_4": 16.24, "d_8": 33.3, "d_16": 56.47, "d_avg": 22.87, "aj": 12.5, "oa": 76.0}, "31-60": {"queries": 2659, "answered": 2656, "visible_queries": 2034, "d_1": 2.11, "d_2": 6.24, "d_4": 16.13, "d_8": 31.91, "d_16": 55.9, "d_avg": 22.46, "aj": 12.34, "oa": 76.57}, "61-120": {"queries": 2120, "answered": 2114, "visible_queries": 1608, "d_1": 2.11, "d_2": 5.6, "d_4": 15.05, "d_8": 29.1, "d_16": 50.56, "d_avg": 20.49, "aj": 11.01, "oa": 76.13}, "121+": {"queries": 2392, "answered": 2389, "visible_queries": 2092, "d_1": 2.63, "d_2": 7.22, "d_4": 15.54, "d_8": 30.5, "d_16": 51.82, "d_avg": 21.54, "aj": 12.58, "oa": 86.83}}, "shards": {"found": 2, "expected": 2, "unfinished": []}}

benchmark_312/qwen3.8-27b-fp8_twostep_single/prompt_twostep_single.txt

@@ -0,0 +1,36 @@
+IMAGE 1 is the reference frame (frame $reference_frame). A single physical surface point is marked with a $marker_color crosshair labeled A. The inset panel shows a $zoom× magnified view of the $area × $area pixel region centered on A.
+
+All remaining images are unmarked target frames:
+$target_lines
+
+Task
+
+For each target frame, locate the same physical surface point marked as A in the reference frame.
+
+Answer in two steps. Step 1, "reference": the position of the crosshair in IMAGE 1. Step 2, "targets": for every id, the position of the same surface point in its target frame. The two steps are different measurements on different images: a target frame has no crosshair, and the camera and the objects have moved.
+
+The apparent position, scale, orientation, and perspective of the object may change between frames. Your goal is to identify the same underlying surface point, defined by the center of the marked reference pixel.
+
+Guidelines
+
+* Track the physical surface point, not its image coordinates.
+* Use all available cues, including:
+    * local texture,
+    * semantic structure,
+    * geometry,
+    * neighboring features,
+    * and object shape.
+
+* Identify the real color of the object; if there is a need, you can utilize the environment to track, otherwise, do not mix the color of the object with the environment
+* Ignore changes in camera viewpoint, scale, lighting, or partial deformation.
+* If the point is occluded, outside the image, or on a surface that has rotated completely out of view, return occluded for that frame rather than estimating its location.
+* $coordinate_rule
+
+Output
+
+* Return exactly one result for each target frame.
+* Copy each target frame ID exactly as provided.
+* Do not omit, reorder, or duplicate any target.
+* Answer with JSON only, no other text:
+{"reference":{"<letter>":[x, y]},"targets":[{"id":"<target id>","point_2d":[x, y],"occluded":<true or false>}]}
+"reference" has one entry per marked point (A), on the 0-1000 scale of IMAGE 1; "targets" has one entry per id, on the 0-1000 scale of its target frame. Use null for point_2d whenever occluded is true.

benchmark_312/qwen3.8-27b-fp8_twostep_single/resolved.yaml

@@ -0,0 +1,76 @@
+name: benchmark_312_twostep_qwen38
+model:
+  model: qwen3.8-27b-fp8
+  hf_id: Qwen/Qwen3.8-27B-FP8
+  call: vlm_track.backends.openai_chat:chat_completion
+  base_url: http://127.0.0.1:8017/v1
+  api_key_env: null
+  timeout: 600
+  convention: norm1000
+  params:
+    max_tokens: 2000
+    temperature: 0.0
+    chat_template_kwargs:
+      enable_thinking: false
+  serve:
+    engine: vllm
+    env: vlm
+    port: 8017
+    min_free_gb: 38
+    max_model_len: 32768
+    max_num_seqs: 32
+    gpu_memory_utilization: 0.88
+    limit_mm_per_prompt:
+      image: 9
+    dtype: bfloat16
+    reasoning_parser: qwen3
+    mm_processor_kwargs:
+      min_pixels: 65536
+      max_pixels: 4194304
+data:
+  name: benchmark_312_variants
+  loader: vlm_track.data.loaders:load_jsonl
+  queries: /p/3dllmselftraining/vlm_3d_tracking_annotator/experiments/benchmark_312/chen_v1_frozen/requests.jsonl
+  truth: /p/3dllmselftraining/vlm_3d_tracking_annotator/experiments/benchmark_312/chen_v1_frozen/truth.jsonl
+  image_root: /p/3dllmselftraining/vlm_3d_tracking_annotator/experiments/benchmark_312/chen_v1_frozen
+  evaluator: vlm_track.evaluation.tapvid:score
+  breakdowns: true
+prompt:
+  file: prompts/multipoint/twostep_single.txt
+  render: vlm_track.prompts.render:tapvid_prompt
+  convention: norm1000
+  text: "IMAGE 1 is the reference frame (frame $reference_frame). A single physical\
+    \ surface point is marked with a $marker_color crosshair labeled A. The inset\
+    \ panel shows a $zoom\xD7 magnified view of the $area \xD7 $area pixel region\
+    \ centered on A.\n\nAll remaining images are unmarked target frames:\n$target_lines\n\
+    \nTask\n\nFor each target frame, locate the same physical surface point marked\
+    \ as A in the reference frame.\n\nAnswer in two steps. Step 1, \"reference\":\
+    \ the position of the crosshair in IMAGE 1. Step 2, \"targets\": for every id,\
+    \ the position of the same surface point in its target frame. The two steps are\
+    \ different measurements on different images: a target frame has no crosshair,\
+    \ and the camera and the objects have moved.\n\nThe apparent position, scale,\
+    \ orientation, and perspective of the object may change between frames. Your goal\
+    \ is to identify the same underlying surface point, defined by the center of the\
+    \ marked reference pixel.\n\nGuidelines\n\n* Track the physical surface point,\
+    \ not its image coordinates.\n* Use all available cues, including:\n    * local\
+    \ texture,\n    * semantic structure,\n    * geometry,\n    * neighboring features,\n\
+    \    * and object shape.\n\n* Identify the real color of the object; if there\
+    \ is a need, you can utilize the environment to track, otherwise, do not mix the\
+    \ color of the object with the environment\n* Ignore changes in camera viewpoint,\
+    \ scale, lighting, or partial deformation.\n* If the point is occluded, outside\
+    \ the image, or on a surface that has rotated completely out of view, return occluded\
+    \ for that frame rather than estimating its location.\n* $coordinate_rule\n\n\
+    Output\n\n* Return exactly one result for each target frame.\n* Copy each target\
+    \ frame ID exactly as provided.\n* Do not omit, reorder, or duplicate any target.\n\
+    * Answer with JSON only, no other text:\n{\"reference\":{\"<letter>\":[x, y]},\"\
+    targets\":[{\"id\":\"<target id>\",\"point_2d\":[x, y],\"occluded\":<true or false>}]}\n\
+    \"reference\" has one entry per marked point (A), on the 0-1000 scale of IMAGE\
+    \ 1; \"targets\" has one entry per id, on the 0-1000 scale of its target frame.\
+    \ Use null for point_2d whenever occluded is true.\n"
+harness:
+  run: vlm_track.harness:single_call
+  detail: high
+  parser: vlm_track.prompts.conventions:norm1000_point_predictions
+concurrency: 8
+limit: null
+shard: null

benchmark_312/qwen3.8-27b-fp8_twostep_single/responses.jsonl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:20aef7a20316f3d7f110358f8bf370d2ce579559d2dda491f11c2a14ab23900b
+size 93021005