Use scalable annotation runtime, shared human-video module and optional Flora captions

README.md

@@ -38,7 +38,8 @@ tags:
 ![LeRobot Annotation Studio — robot video to language](assets/studio-cover.png)
 
 Enter a LeRobot v3 dataset, choose annotations, review an estimate, then sign in
-with Hugging Face to run a CPU Job using the selected vision model through Inference Providers.
+with Hugging Face to run a Job using the selected annotation model. General vision
+models use a CPU Job and Inference Providers; Flora uses an H200 GPU Job.
 Jobs and inference are billed to the signed-in user's personal account.
 The Space itself uses free CPU hardware and has no shared HF access token.
 
@@ -62,8 +63,9 @@ A Studio adapter preserves the quoted first-frame-only workflow, full-resolution
 image edits, spending/time guards and the media viewer. End-frame editing remains
 available in the CLI but is not included in the Studio's single-edit quote.
 
-In **Model → Subtask captions**, choose **NVIDIA Flora-Captioner** to relabel
-the intervals generated by the selected vision model. Flora preserves those
+In the **Model** selector, choose **NVIDIA Flora-Captioner** to relabel
+the intervals generated by its Qwen3.8 27B helper. The helper is shown beneath
+the selector and in the review, with its provider rates. Flora preserves those
 boundaries; the relabeled subtasks then feed plans, memory and human videos.
 It uses NVIDIA's pinned native frame planner, video processor, mixed BF16/FP32
 loader and greedy decoder with BEFORE/TARGET/AFTER context. The input uses the
@@ -72,12 +74,12 @@ there is no silent fallback to the original VLM text.
 
 Flora has no hosted Inference Provider route. Selecting it runs the worker on an
 H200 CUDA Job; the quote reserves the entire GPU timeout, including installation,
-weight downloads and startup. Other annotations still use the selected VLM through
+weight downloads and startup. Other annotations use the helper VLM through
 Inference Providers. The default keeps the CPU workflow.
 [Flora's model card](https://huggingface.co/nvidia/Flora-Captioner) limits use to
 noncommercial research. Code and weight revisions are pinned in `flora_captioner.py`
-and each run manifest. This option is experimental; a paired 25-episode ABC
-captioning comparison is being evaluated separately from boundary detection.
+and each run manifest. This option is experimental. `benchmark_flora.py` supports
+a paired 25-episode ABC captioning comparison separately from boundary detection.
 
 The default **tuned Qwen3.8** preset uses the action-phase prompt developed on
 ABC, timestamped contact sheets at 2 fps, 1,200 frames per request, and up to