Trim model card, point to GitHub for usage

README.md

@@ -17,14 +17,10 @@ pipeline_tag: text-classification
 
 # ORCA — Llama-3.2-3B-Instruct (Multinomial, seed 99)
 
-ORCA (**O**pen-ended **R**esponse **C**orrectness **A**ssessment) is a scoring model for audio question-answering tasks. Given a question, a reference answer, a candidate answer, and an LLM-generated rationale, ORCA outputs a correctness score and an uncertainty estimate.
+ORCA (**O**pen-ended **R**esponse **C**orrectness **A**ssessment) scores the correctness of open-ended audio QA responses. Given a question, reference answer, candidate answer, and an LLM-generated rationale, it outputs a correctness score in [0, 1] and an uncertainty estimate.
 
-This checkpoint uses **Meta Llama-3.2-3B-Instruct** as the backbone with a LoRA adapter (rank 128) and a linear scoring head trained with the **multinomial log-likelihood** loss on a 5-point Likert scale.
-
-**Paper:** [ORCA: Open-ended Response Correctness Assessment for Audio Question Answering](https://arxiv.org/abs/2512.09066)
-Accepted to *Transactions of the Association for Computational Linguistics (TACL), 2026*.
-
-**Code:** [github.com/BUT-FIT/orca](https://github.com/BUT-FIT/orca)
+**Paper:** [ORCA: Open-ended Response Correctness Assessment for Audio Question Answering](https://arxiv.org/abs/2512.09066) — accepted to *TACL 2026*  
+**Code & usage:** [github.com/BUTSpeechFIT/ORCA](https://github.com/BUTSpeechFIT/ORCA)  
 **Training data:** [BUT-FIT/orca-audio-qa-annotations](https://huggingface.co/datasets/BUT-FIT/orca-audio-qa-annotations)
 
 ## Model details
@@ -33,84 +29,20 @@ Accepted to *Transactions of the Association for Computational Linguistics (TACL
 |---|---|
 | Base model | `meta-llama/Llama-3.2-3B-Instruct` |
 | LoRA rank / alpha | 128 / 128 |
-| LoRA target modules | q, k, v, o, gate, up, down projections |
 | Loss function | Multinomial log-likelihood (5-class Likert) |
 | Training seed | 99 |
-| Training curriculum | Stage 1 (synthetic) → Stage 2 (LLM-judge) → Stage 3 (human annotations) |
+| Training curriculum | Stage 1 (synthetic) → Stage 2 (LLM-judge) → Stage 3 (human) |
 | Precision | bfloat16 |
 
-## Repository structure
-
-```
-model/
-├── config.yaml               # ORCA config (score_type: multinomial)
-├── model_minus_lm.pt         # Linear scoring head weights
-└── lm/
-    ├── adapter_config.json   # LoRA adapter config
-    └── adapter_model.safetensors
-tokenizer/                    # Saved tokenizer (same as base model)
-```
-
-## Usage
-
-### Install
+## Quick start
 
 ```bash
-pip install git+https://github.com/BUT-FIT/orca.git
-# or, after cloning:
-pip install -e .
-```
-
-### Download this model
-
-```bash
-pip install huggingface_hub
+pip install git+https://github.com/BUTSpeechFIT/ORCA.git
 hf download BUT-FIT/orca-llama-3.2-3b-it-multinomial --local-dir orca-llama-3b
+orca-infer --model_path orca-llama-3b/model --data_jsonl your_data.jsonl --output_dir results/
 ```
 
-### Prepare input data
-
-Input is a JSONL file, one item per line:
-
-```json
-{"id": "example_1", "question": "How many instruments are played in this recording?", "reference": "Three instruments: piano, violin, and cello.", "candidate": "I hear a piano and violin being played.", "rationale": "The candidate identifies two of the three instruments but misses the cello.", "ratings": []}
-```
-
-Fields:
-- `question`: the audio QA question
-- `reference`: the ground-truth reference answer
-- `candidate`: the model response to evaluate
-- `rationale`: an LLM-generated explanation of why the candidate is correct or incorrect
-- `ratings`: list of human ratings (integers 1–5); leave as `[]` for unlabeled data
-
-### Run inference
-
-```bash
-orca-infer \
-    --model_path orca-llama-3b/model \
-    --data_jsonl your_data.jsonl \
-    --output_dir results/
-```
-
-The output `results/your_data/final_result.jsonl` contains one row per input with added fields:
-- `rating_orca`: correctness score in [0, 1]
-- `variance_orca`: uncertainty estimate
-- `params`: raw distribution parameters
-
-If your JSONL contains `ratings`, evaluation metrics (Spearman, Kendall, MAE) are printed and saved.
-
-### Download evaluation splits (seed 99)
-
-```bash
-hf download BUT-FIT/orca-audio-qa-annotations \
-    --type dataset --local-dir ./data \
-    --include "stage3_human/seed_99/*"
-
-orca-infer \
-    --model_path orca-llama-3b/model \
-    --data_jsonl data/stage3_human/seed_99/test.jsonl \
-    --output_dir results/
-```
+See the [repository](https://github.com/BUTSpeechFIT/ORCA) for full usage, evaluation scripts, and the `download_and_infer.py` convenience script.
 
 ## Citation
 
@@ -129,4 +61,4 @@ orca-infer \
 
 ## License
 
-MIT License. See the [repository LICENSE](https://github.com/BUT-FIT/orca/blob/main/LICENSE) for details.
+MIT License. See the [repository LICENSE](https://github.com/BUTSpeechFIT/ORCA/blob/main/LICENSE) for details.