Upload folder using huggingface_hub

app.py

@@ -6,6 +6,7 @@ import spaces  # MUST come before torch / any CUDA-touching import
 import torch
 import gradio as gr
 import numpy as np
+import librosa
 from transformers import (
     Qwen2_5OmniForConditionalGeneration,
     Qwen2_5OmniProcessor,
@@ -20,6 +21,11 @@ model = Qwen2_5OmniForConditionalGeneration.from_pretrained(
     attn_implementation="sdpa",
 ).to("cuda").eval()
 
+THINK_OPEN = "<think>"
+THINK_CLOSE = "</think>"
+ANSWER_OPEN = "<answer>"
+ANSWER_CLOSE = "</answer>"
+
 
 @spaces.GPU(duration=120)
 def answer_audio_question(
@@ -39,6 +45,8 @@ def answer_audio_question(
         temperature: Sampling temperature (0.0 = greedy).
         enable_thinking: If True, the model reasons step-by-step before answering.
     """
+    import re
+
     if audio_path is None:
         return "Please upload an audio file.", ""
     if not question.strip():
@@ -49,14 +57,26 @@ def answer_audio_question(
         system_content = (
             "You are an expert audio understanding assistant. "
             "Listen carefully and answer questions about the audio. "
-            "Always think step by step inside <think> tags, "
-            "then give the final answer inside <answer> tags."
+            "Always think step by step inside "
+            + THINK_OPEN
+            + " tags, "
+            "then give the final answer inside "
+            + ANSWER_OPEN
+            + " tags."
         )
         user_text = (
             f"Listen to the audio carefully and answer the following question.\n\n"
             f"Question: {question}\n\n"
-            "First, reason step by step inside tags.\n"
-            "Then output your final answer inside <answer> ... </answer> tags."
+            "First, reason step by step inside "
+            + THINK_OPEN
+            + " ... "
+            + THINK_CLOSE
+            + " tags.\n"
+            "Then output your final answer inside "
+            + ANSWER_OPEN
+            + " ... "
+            + ANSWER_CLOSE
+            + " tags."
         )
     else:
         system_content = (
@@ -84,13 +104,12 @@ def answer_audio_question(
         messages, tokenize=False, add_generation_prompt=True
     )
 
-    # Load audio from file
-    from pathlib import Path
+    # Load audio as numpy array (resampled to 16kHz for Whisper feature extractor)
+    audio_data, sr = librosa.load(audio_path, sr=16000, mono=True)
 
-    audio_file = Path(audio_path)
     inputs = processor(
         text=text,
-        audios=[str(audio_file)],
+        audio=audio_data,
         return_tensors="pt",
         padding=True,
     ).to("cuda").to(model.dtype)
@@ -109,26 +128,25 @@ def answer_audio_question(
     response = processor.decode(generated_ids, skip_special_tokens=True)
 
     # Parse thinking and answer
-    import re
+    think_pattern = re.escape(THINK_OPEN) + r"\s*(.*?)\s*" + re.escape(THINK_CLOSE)
+    answer_pattern = re.escape(ANSWER_OPEN) + r"\s*(.*?)\s*" + re.escape(ANSWER_CLOSE)
 
-    think_match = re.search(
-        r"<think>\s*(.*?)\s*</think>", response, flags=re.DOTALL | re.IGNORECASE
-    )
-    answer_match = re.search(
-        r"<answer>\s*(.*?)\s*</answer>", response, flags=re.DOTALL | re.IGNORECASE
-    )
+    think_match = re.search(think_pattern, response, flags=re.DOTALL | re.IGNORECASE)
+    answer_match = re.search(answer_pattern, response, flags=re.DOTALL | re.IGNORECASE)
 
     thinking_text = think_match.group(1).strip() if think_match else ""
     answer_text = answer_match.group(1).strip() if answer_match else ""
 
     # If no tags found, return the full response as the answer
-    if not answer_text:
+    if not answer_text and not thinking_text:
         answer_text = response.strip()
         thinking_text = ""
+    elif not answer_text:
+        answer_text = response.strip()
 
     # Format nicely
     if thinking_text:
-        formatted_thinking = f"💭 **Reasoning:**\n{thinking_text}"
+        formatted_thinking = f"**Reasoning:**\n{thinking_text}"
     else:
         formatted_thinking = ""
 
@@ -143,10 +161,10 @@ CSS = """
 with gr.Blocks() as demo:
     with gr.Column(elem_id="col-container"):
         gr.Markdown(
-            "# 🎧 AudioRubrics: Audio Reasoning with Evolving Rubric Rewards\n"
+            "# AudioRubrics: Audio Reasoning with Evolving Rubric Rewards\n"
             "Upload an audio clip and ask a question about it. The model reasons step-by-step "
             "about what it hears.\n\n"
-            "Based on [Reinforcement Learning with Evolving Rubrics as Rewards for Audio Reasoning](https://huggingface.co/papers/2608.02831) — "
+            "Based on [Reinforcement Learning with Evolving Rubrics as Rewards for Audio Reasoning](https://huggingface.co/papers/2608.02831) | "
             "a Qwen2.5-Omni-7B model fine-tuned with GRPO using self-evolving, audio-grounded rubric rewards."
         )
 

requirements.txt

@@ -3,4 +3,5 @@ accelerate
 sentencepiece
 soundfile
 numpy
-torchvision
+torchvision
+librosa