Update README.md

README.md

@@ -11,48 +11,75 @@ tags:
 library_name: transformers
 base_model: openai/whisper-tiny
 pipeline_tag: automatic-speech-recognition
+datasets:
+- reazon-research/reazonspeech
+metrics:
+- cer
+- loss
+model-index:
+- name: whisper-tiny-ja-lora
+  results:
+  - task:
+      type: automatic-speech-recognition
+      name: Automatic Speech Recognition
+    dataset:
+      name: japanese-asr/ja_asr.reazonspeech_test
+      type: japanese-asr/ja_asr.reazonspeech_test
+      split: test
+    metrics:
+    - type: cer
+      name: Character Error Rate (CER)
+      value: 0.52497
+    - type: loss
+      name: Eval Loss
+      value: 1.17656
 ---
 
 # Whisper Tiny JA LoRA (ReazonSpeech)
 
 LoRA adapter fine-tuned from `openai/whisper-tiny` for Japanese ASR.
 
-- Base model: `openai/whisper-tiny`
-- Training method: LoRA (`q_proj`, `v_proj`)
-- Dataset: `reazon-research/reazonspeech` (gated)
-- Language: Japanese (`ja`)
-
 ## Model Type
 
 This repository contains **LoRA adapter weights only**.  
-To run inference, load this adapter on top of `openai/whisper-tiny`.
+Use it on top of `openai/whisper-tiny`.
+
+- Base model: `openai/whisper-tiny`
+- Language: Japanese (`ja`)
+- Training method: LoRA (`q_proj`, `v_proj`)
+- Dataset: `reazon-research/reazonspeech` (gated)
 
 ## Training Setup
 
-- Epochs: `3`
+- Epochs (configured): `3`
 - Learning rate: `1e-5`
+- Batch size: `32`
 - LoRA r / alpha / dropout: `16 / 32 / 0.05`
-- Batch size: `32` (or your actual value)
 - Framework: `transformers`, `peft`
+- Runtime: Kaggle GPU P100
 
-## Intended Use
+## Evaluation (Latest W&B Run)
 
-- Japanese speech-to-text transcription
-- Lightweight adaptation with small trainable parameter count
+- `eval/cer`: **0.52497** (52.50%)
+- `eval/loss`: **1.17656**
+- `eval/runtime`: **162.422 s**
+- `eval/samples_per_second`: **12.314**
+- `eval/steps_per_second`: **0.77**
+- `train/global_step`: **3000**
+- `train/epoch`: **1.54719**
 
-## Limitations
+> Note: WER was not logged in this run.
 
-- Performance may degrade on domain/audio conditions not covered by ReazonSpeech
-- Not evaluated for safety-critical use cases
-- Dataset access requires accepted terms on Hugging Face
+## Intended Use
 
-## Evaluation
+- Japanese speech-to-text transcription
+- Lightweight adapter training and deployment
 
-Fill with your real metrics:
+## Limitations
 
-- CER: `TODO`
-- WER: `TODO`
-- Eval split: `TODO`
+- Quality depends on domain/audio condition match with training data
+- Not validated for safety-critical production use
+- Requires accepted access to gated dataset when reproducing training
 
 ## Load Adapter
 
@@ -61,7 +88,7 @@ from transformers import WhisperForConditionalGeneration, WhisperProcessor
 from peft import PeftModel
 
 base_model_id = "openai/whisper-tiny"
-adapter_id = "dungca/whisper-tiny-ja-lora"  # replace if needed
+adapter_id = "dungca/whisper-tiny-ja-lora"
 
 processor = WhisperProcessor.from_pretrained(base_model_id)
 base_model = WhisperForConditionalGeneration.from_pretrained(base_model_id)