Fix generation config (persistable prompt) + CV25 dataset reference

README.md

@@ -5,7 +5,7 @@ library_name: transformers
 pipeline_tag: automatic-speech-recognition
 base_model: openai/whisper-large-v3
 tags: [automatic-speech-recognition, whisper, mongolian, buzzasr]
-datasets: [google/fleurs, mozilla-foundation/common_voice_17_0]
+datasets: [google/fleurs]
 metrics: [cer, wer]
 ---
 
@@ -15,18 +15,18 @@ A monolingual automatic speech recognition model for **Mongolian**, fine-tuned f
 [openai/whisper-large-v3](https://huggingface.co/openai/whisper-large-v3). Part of **BuzzASR**,
 a suite of 102 language-specialized ASR models (Findings of EMNLP 2026).
 
-This model uses **full fine-tuning** (native per-language tokenizer replacement + text multitask fine-tuning).
-It reaches **state-of-the-art CER among open-source systems** on Mongolian.
+This model uses **full fine-tuning** (native per-language tokenizer replacement + text multitask fine-tuning),
+and reaches **state-of-the-art CER among open-source systems** on Mongolian.
 
 ## Results (normalized CER / WER, %)
 
 | Test set | CER | WER | Whisper-large-v3 (zero-shot) CER |
 |---|---|---|---|
 | FLEURS | 9.15 | 24.69 | 34.42 |
-| CommonVoice v25 | 1.00 | 2.14 | 40.14 |
+| Common Voice 25 | 1.00 | 2.14 | 40.14 |
 | **Combined** | **5.21** | **13.34** | **38.23** |
 
-That is a **~7.3× CER reduction** over Whisper zero-shot on the combined test set.
+~7.3x CER reduction over Whisper zero-shot on the combined test set.
 
 ## Usage
 
@@ -42,13 +42,14 @@ feats = proc(wav[0], sampling_rate=16000, return_tensors="pt").input_features.to
 ids = model.generate(feats, num_beams=1, no_repeat_ngram_size=3, repetition_penalty=1.2)
 print(proc.batch_decode(ids, skip_special_tokens=True)[0])
 ```
-The language/task prompt is baked into the model's generation config, so no `language=` argument is needed.
+The language/task prompt is baked into the generation config, so no `language=` argument is needed.
 
 ## Training data
-FLEURS + CommonVoice v25 (Mongolian), capped per the paper. Text-only data from the fish-food corpus.
+[FLEURS](https://huggingface.co/datasets/google/fleurs) + **Common Voice Corpus 25.0** (Mozilla, March 2025;
+https://commonvoice.mozilla.org/en/datasets), capped per the paper. Text-only data from the fish-food corpus.
 
 ## Limitations
-Monolingual — intended for Mongolian only. Evaluated on FLEURS/CommonVoice test splits; other domains/dialects may differ.
+Monolingual (Mongolian only). Evaluated on FLEURS / Common Voice test splits; other domains or dialects may differ.
 
 ## Citation
-See the project page: https://lemn-lab.github.io/buzzasr-docs/
+Project page: https://lemn-lab.github.io/buzzasr-docs/

generation_config.json

@@ -7,22 +7,18 @@
   "bos_token_id": 47472,
   "decoder_start_token_id": 47473,
   "eos_token_id": 47472,
-  "forced_decoder_ids": [
-    [
-      1,
-      47529
-    ],
-    [
-      2,
-      47575
-    ],
-    [
-      3,
-      47579
-    ]
-  ],
+  "lang_to_id": {
+    "<|mn|>": 47529
+  },
+  "language": "mn",
   "max_length": 448,
+  "no_timestamps_token_id": 47579,
   "pad_token_id": 47472,
   "suppress_tokens": [],
+  "task": "transcribe",
+  "task_to_id": {
+    "transcribe": 47575,
+    "translate": 47574
+  },
   "transformers_version": "4.57.6"
 }