Update README.md

README.md

@@ -14,7 +14,7 @@ This dataset releases **data selection/preparation and the dialogue test set**.
 
 | Path | Contents |
 | --- | --- |
-| `testset/` | 67 duplex and 71 semantic clips, frozen `judge_context.json`, and `protocol.json` |
+| `testset/` | ELLSA 2 dialogue gesture benchmark, 67 duplex and 71 semantic clips |
 | `manifests/train_all_resolved.json` | Full source pool: 121,054 aligned target/partner recordings, with portable paths |
 | `manifests/train_selected81.json` | Selected inventory of 1,318 recordings from 81 speakers |
 | `manifests/selected_speakers.json` | Selected speaker IDs |
@@ -74,9 +74,7 @@ conda activate ellsa2_kaolrm
 python -m pip install -r preprocess/requirements-kaolrm.txt
 ```
 
-The checkpoint package includes `config.json`, `model.safetensors`, and their weight license under `../checkpoints/external_assets/kaolrm-mono/`. Their source is the [official KaoLRM mono release](https://github.com/CyberAgentAILab/KaoLRM/releases/tag/mono); see the [asset guide](../checkpoints/README.md#external-assets-sources-and-download-guide) for attribution and independent downloads. You can also pass another checkpoint directory. The checkpoint contains its DINO encoder.
-
-Resolve the selected inventory against your download, then extract labels. The existing `flame_file` paths in the input describe expected labels; those files do not need to exist before extraction. Each row needs `pose_file`; the video defaults to the same basename with `.mp4`, or can be specified as `video_file`. Pose NPZs need `boxes_and_keypoints:keypoints` and `smplh:is_valid`; face landmarks occupy indices 23–90 in the original detector layout.
+Resolve the selected inventory against your download, then extract labels.
 
 ```bash
 python preprocess/resolve_paths.py \
@@ -92,9 +90,7 @@ python preprocess/extract_kaolrm.py \
   --device cuda:0
 ```
 
-The extractor writes one `<pose_stem>_flame.npz` per recording into `outputs/flame_raw/`, and an absolute-path training manifest into `outputs/train_raw.json`. Use `--max-videos 2 --max-frames 96` for a small check, with a separate output directory. Omit both caps for the full inventory. Sharded processes must use distinct output manifests. Existing outputs require explicit `--overwrite`; duplicate pose stems fail.
-
-Crop trajectories use the original 0.35 face margin, confidence threshold 0.05, 15/21-frame center/size smoothing and 0.10 coverage margin. Missing face measurements are interpolated for crop placement on valid body frames. Invalid body frames remain invalid; frames with no usable crop are zero-filled and excluded by `valid_mask`. Crop smoothing uses the complete source trajectory even when `--max-frames` limits the output. Facial parameters remain raw, with per-frame shape and no parameter smoothing. The NPZ stores 50 predicted expression coefficients padded to 100, six pose parameters, six zero eye parameters, and a 112D `motioncode`; the training loader selects the 108D representation used by the RVQ.
+The extractor writes one `<pose_stem>_flame.npz` per recording into `outputs/flame_raw/`, and an absolute-path training manifest into `outputs/train_raw.json`. Use `--max-videos 2 --max-frames 96` for a small check, with a separate output directory.
 
 ### Link source data and extracted labels
 
@@ -106,13 +102,13 @@ python preprocess/link_training_data.py \
   --flame-root outputs/flame_raw
 ```
 
-This creates `source/seamless_interaction/` under the data package. Audio, poses and transcripts link to the downloaded dataset; `flame_file` links to the newly extracted flat output directory. Every referenced file must exist. The command can be rerun with the same inputs and does not copy the recordings. If you already have compatible raw `_flame.npz` files beside the source poses, omit `--flame-root` to use those files.
+This creates `source/seamless_interaction/` under the data package. Audio, poses and transcripts link to the downloaded dataset; `flame_file` links to the newly extracted flat output directory.
 
 Training can now use the included `manifests/train_selected81.json` unchanged, as described in the [training guide](../code/README.md#training).
 
 ## Speaker selection
 
-The included `manifests/train_all_resolved.json` supplies the **121,054-recording source pool** used for the frozen 3,401-speaker quality statistics. It already contains the paired target/partner filenames, so no source-pool manifest needs to be assembled. Its paths are portable `data/source/seamless_interaction/...` references; `--dataset-root` maps them to your downloaded files. Absolute paths from the original machine are not required. Scoring reads poses and transcripts and does not require FLAME extraction.
+The included `manifests/train_all_resolved.json` supplies the **121,054-recording source pool** used for the speaker quality statistics. Its paths are portable `data/source/seamless_interaction/...` references; `--dataset-root` maps them to your downloaded files.
 
 ### 1. Score 60-second windows
 
@@ -126,7 +122,7 @@ python preprocess/build_stageA_quality_data.py \
   --target-sec 60 --stride-sec 30 --num-workers 4
 ```
 
-The default scale uses source-pool quantiles frozen in `selection_protocol.json`. Use `--recompute-normalization` only when fitting a replacement full-pool scale. The output retains full recordings from selected speakers; it does not trim training into individual passing windows. If you already have an absolute-path `train_all_resolved.json`, pass it directly and omit `--dataset-root`.
+The default scale uses source-pool quantiles frozen in `selection_protocol.json`. Use `--recompute-normalization` only when fitting a replacement full-pool scale. If you already have an absolute-path `train_all_resolved.json`, pass it directly and omit `--dataset-root`.
 
 | Quality component | Weight |
 | --- | ---: |
@@ -152,9 +148,7 @@ python preprocess/build_stageA_select_spk.py \
   --min-windows 1 --min-passed 1
 ```
 
-These thresholds retain speakers with pass rate ≥ 0.70, mean quality ≥ 0.90, mean speech-motion compatibility ≥ 0.90, outlier rate ≤ 0.01, and at least one assessed/passing window. The legacy CSV `selected` column does not override this rule. The replay produces **exactly the 81 supplied speaker IDs and the same 1,318 target/partner recordings**. Facial labels are attached after selection in `train_selected81.json`; speaker selection itself does not depend on them.
-
-For newly measured data, replace the CSV input with `outputs/speaker_quality_new.csv`. To use another source pool, supply an equivalent paired manifest with `pose_file`, `audio_file`, `audio_file_listen`, `word_file`, and optional `word_file_listen`. Both audio streams must share their timeline.
+These thresholds retain speakers with pass rate ≥ 0.70, mean quality ≥ 0.90, mean speech-motion compatibility ≥ 0.90, outlier rate ≤ 0.01, and at least one assessed/passing window.
 
 ## License and distribution