641071660J commited on
Commit
33bea16
·
verified ·
1 Parent(s): 9873d2b

Document the current auxiliary annotation record counts

Browse files
Files changed (1) hide show
  1. README.md +20 -1
README.md CHANGED
@@ -39,7 +39,7 @@ Each example references two images: an ISO reference and the FreeCAD screenshot
39
  - `annotations/train_final.jsonl.gz`, `annotations/test_final.jsonl.gz`: byte-preserved canonical annotations after decompression.
40
  - `images/iso_images.tar`: one reference image per part.
41
  - `images/train-*.tar`, `images/test-*.tar`: GUI screenshots, sharded by part and split, retaining original relative paths.
42
- - `subsets/`: existing correction-SFT, GRPO, and evaluation subset annotations for reproducing earlier experiments. These are subsets or derived training records, **not additional independent splits**. In particular, `test_50.jsonl` was used during development; the accompanying 364-part manifest describes the remainder of the canonical test split.
43
  - `metadata/source_manifest.json`: original annotation hashes, split statistics, and packaging provenance.
44
  - `metadata/image_manifests/*.jsonl.gz`: SHA-256, size, and original path for every packaged image.
45
  - `metadata/release_manifest.json`, `SHA256SUMS`: release payload verification.
@@ -69,3 +69,22 @@ GUI_upload/
69
  Resolve each annotation's `images` entries relative to `GUI_upload`. The two-image order remains ISO reference first, current GUI screenshot second. `messages`, `present_task`, `task_progress`, label columns, bounding boxes, and source identifiers are retained exactly as in the canonical JSONL files.
70
 
71
  This archive-based release is intended for full local reconstruction; the repository viewer is not a substitute for extracting the image archives. Original annotation records remain the authoritative schema.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
39
  - `annotations/train_final.jsonl.gz`, `annotations/test_final.jsonl.gz`: byte-preserved canonical annotations after decompression.
40
  - `images/iso_images.tar`: one reference image per part.
41
  - `images/train-*.tar`, `images/test-*.tar`: GUI screenshots, sharded by part and split, retaining original relative paths.
42
+ - `subsets/`: snapshots of existing correction-SFT, GRPO, and evaluation helper annotations. These are subsets or derived training records, **not additional independent splits**. In particular, `test_50.jsonl` was used during development; the accompanying 364-part manifest describes the remainder of the canonical test split.
43
  - `metadata/source_manifest.json`: original annotation hashes, split statistics, and packaging provenance.
44
  - `metadata/image_manifests/*.jsonl.gz`: SHA-256, size, and original path for every packaged image.
45
  - `metadata/release_manifest.json`, `SHA256SUMS`: release payload verification.
 
69
  Resolve each annotation's `images` entries relative to `GUI_upload`. The two-image order remains ISO reference first, current GUI screenshot second. `messages`, `present_task`, `task_progress`, label columns, bounding boxes, and source identifiers are retained exactly as in the canonical JSONL files.
70
 
71
  This archive-based release is intended for full local reconstruction; the repository viewer is not a substitute for extracting the image archives. Original annotation records remain the authoritative schema.
72
+
73
+ ## Auxiliary annotation inventory
74
+
75
+ The following counts describe the files currently present on the source server. The canonical training and test sets are the two files under `annotations/`; auxiliary filenames do not establish dataset size or correspondence to a final model checkpoint. For example, the file whose name includes `gt8192` currently contains 3 records, and `train_action_grpo_correction_balanced_v2.jsonl` contains 8 records. They are preserved as found.
76
+
77
+ | Auxiliary JSONL source | Records |
78
+ |---|---:|
79
+ | `test_50.jsonl` | 3,352 |
80
+ | `train_action_correction_sft_v1.jsonl` | 590 |
81
+ | `train_action_correction_sft_v1_gt8192.jsonl` | 3 |
82
+ | `train_action_grpo.jsonl` | 11,894 |
83
+ | `train_action_grpo_click_calibrated.jsonl` | 12,000 |
84
+ | `train_action_grpo_correction_balanced_v2.jsonl` | 8 |
85
+ | `train_action_grpo_correction_balanced_v2.merged_groups.jsonl` | 54 |
86
+ | `train_action_grpo_correction_balanced_v2.merged_reference.jsonl` | 54 |
87
+ | `train_action_grpo_targeted256_targeted480_correction75_k8_v1_20260902.jsonl` | 256 |
88
+ | `val_action_grpo.jsonl` | 104 |
89
+
90
+ The accompanying JSON configuration and split-manifest files are also included under `subsets/`.