Streamline evaluation guides and resource navigation

README.md

@@ -2,7 +2,7 @@
 pretty_name: PsycheMME
 license: other
 license_name: source-specific
-license_link: https://huggingface.co/datasets/psychemme/PsycheMME/blob/v1.0-paper.1/licenses/THIRD_PARTY_NOTICES.md
+license_link: https://huggingface.co/datasets/psychemme/PsycheMME/blob/v1.0.2/licenses/THIRD_PARTY_NOTICES.md
 language:
   - en
 task_categories:
@@ -36,15 +36,15 @@ pip install datasets huggingface_hub
 ```python
 from datasets import load_dataset
 
-questions = load_dataset("psychemme/PsycheMME", "questions", revision="v1.0-paper.1", split="test")
-asr = load_dataset("psychemme/PsycheMME", "asr", revision="v1.0-paper.1", split="test")
+questions = load_dataset("psychemme/PsycheMME", "questions", revision="v1.0.2", split="test")
+asr = load_dataset("psychemme/PsycheMME", "asr", revision="v1.0.2", split="test")
 ```
 
 Download all benchmark files, including the videos:
 
 ```bash
 hf download psychemme/PsycheMME --repo-type dataset \
-  --revision v1.0-paper.1 --local-dir ./PsycheMME
+  --revision v1.0.2 --local-dir ./PsycheMME
 ```
 
 ## Benchmark
@@ -68,11 +68,23 @@ The 630 videos occupy about 27.6 GB, including 61 clips at 4K UHD. Dimensions, f
 
 The episodes come from MERCI+, CASTLE, EasyCom, IndEgo, Ego4D, EgoCom, YouTube Vlog and YouTube Interview. Video, audio and source-derived text follow their respective [source terms](licenses/THIRD_PARTY_NOTICES.md).
 
-## Evaluation and Paper Results
+## Evaluate Your Model
 
-Evaluation tools are maintained in [PsycheMME on GitHub](https://github.com/psychemme/PsycheMME). The paper's 19,000 scoring records, result tables and score replay inputs are published separately in [PsycheMME-logs](https://huggingface.co/datasets/psychemme/PsycheMME-logs).
+Install the [evaluation package](https://github.com/psychemme/PsycheMME#-installation), configure your model and Judge endpoints in `.env`, then run:
 
-See [Data Format](docs/DATA_FORMAT.md) for fields and [the project page](https://psychemme.github.io/) for methods, results and examples.
+```bash
+bash scripts/evaluate.sh
+```
+
+For a ten-question first run with only the required videos:
+
+```bash
+RUN_DIR=runs/sample bash scripts/evaluate.sh 10
+```
+
+Repeat the same command to continue an interrupted evaluation. See [Quick Start](https://github.com/psychemme/PsycheMME/blob/main/docs/QUICKSTART.md) for setup and outputs, and [Data Format](docs/DATA_FORMAT.md) for fields.
+
+Paper scoring factors and tables are available in [PsycheMME-logs](https://huggingface.co/datasets/psychemme/PsycheMME-logs).
 
 ## Citation
 

docs/DATA_FORMAT.md

@@ -57,4 +57,3 @@ Paper scoring records and their compact ID mapping are in the separate [PsycheMM
 
 Forty Memory questions include a nonempty `memory_context`, passed unchanged to the model and the reference-memory Judge. The paper's explicit-history ablation uses these 40 items. A null value means no separate history text is supplied; the video and ASR remain model inputs.
 
-Fourteen questions are marked as history-dependent but have null supplied context in the paper source. These records are preserved in the paper data. `memory_dependency` and reference evidence are not substitutes for model-facing context.