Update app.py

app.py

@@ -18,15 +18,15 @@ from jiwer import cer
 
 # ---------------- Storage & Paths ----------------
 os.makedirs("data", exist_ok=True)
-FEEDBACK_PATH = "data/feedback.jsonl" # raw feedback log (per sample)
-MEMORY_RULES_PATH = "data/memory_rules.json" # compiled post-correction rules
-GRPO_EXPORT_PATH = "data/grpo_prefs.jsonl" # preference pairs for GRPO
-CSV_EXPORT_PATH = "data/feedback.csv" # optional tabular export
+FEEDBACK_PATH = "data/feedback.jsonl"  # raw feedback log (per sample)
+MEMORY_RULES_PATH = "data/memory_rules.json"  # compiled post-correction rules
+GRPO_EXPORT_PATH = "data/grpo_prefs.jsonl"  # preference pairs for GRPO
+CSV_EXPORT_PATH = "data/feedback.csv"  # optional tabular export
 
 # ---------------- Models ----------------
 MODEL_PATHS = {
-    "Model 1 (Complex handwrittings )": ("prithivMLmods/Qwen2.5-VL-7B-Abliterated-Caption-it", Qwen2_5_VLForConditionalGeneration),
-    "Model 2 (simple and scanned handwritting )": ("nanonets/Nanonets-OCR-s", Qwen2_5_VLForConditionalGeneration),
+    "Model 1 (Complex handwritings)": ("prithivMLmods/Qwen2.5-VL-7B-Abliterated-Caption-it", Qwen2_5_VLForConditionalGeneration),
+    "Model 2 (simple and scanned handwriting)": ("nanonets/Nanonets-OCR-s", Qwen2_5_VLForConditionalGeneration),
 }
 # Model 3 removed to conserve memory.
 
@@ -161,7 +161,7 @@ def _compile_rules_from_feedback(min_count: int = 2, max_phrase_len: int = 40):
                 row = json.loads(line)
             except Exception:
                 continue
-            if row.get("reward", 0) < 1: # only learn from thumbs-up or explicit 'accepted_correction'
+            if row.get("reward", 0) < 1:  # only learn from thumbs-up or explicit 'accepted_correction'
                 continue
             pred = _safe_text(row.get("prediction", ""))
             corr = _safe_text(row.get("correction", "")) or _safe_text(row.get("ground_truth", ""))
@@ -285,14 +285,12 @@ def _export_csv():
             w.writerow(flat)
     return CSV_EXPORT_PATH
 
-# ------------------- MODIFIED -------------------
 def save_feedback(image: Image.Image, model_choice: str, prompt: str,
                   prediction: str, correction: str, ground_truth: str, reward: int):
     """
     reward: 1 = good/accepted, 0 = neutral, -1 = bad
     """
     if image is None:
-        # Bug Fix: Return a single string, not a tuple
         return "Please provide the image again to link feedback."
     if not prediction and not correction and not ground_truth:
         return "Nothing to save."
@@ -322,7 +320,6 @@ def save_feedback(image: Image.Image, model_choice: str, prompt: str,
     }
     _append_jsonl(FEEDBACK_PATH, row)
     return f"✅ Feedback saved (reward={reward})."
-# ------------------------------------------------
 
 def compile_memory_rules():
     _compile_rules_from_feedback(min_count=2, max_phrase_len=60)
@@ -360,7 +357,6 @@ def export_grpo_preferences():
                     count += 1
     return f"✅ Exported {count} GRPO preference pairs to {GRPO_EXPORT_PATH}."
 
-# ------------------- NEW -------------------
 def get_grpo_file():
     if os.path.exists(GRPO_EXPORT_PATH):
         return GRPO_EXPORT_PATH
@@ -371,7 +367,6 @@ def get_csv_file():
     if os.path.exists(CSV_EXPORT_PATH):
         return CSV_EXPORT_PATH
     return None
-# -------------------------------------------
 
 # ---------------- Evaluation Orchestration ----------------
 @spaces.GPU
@@ -448,7 +443,7 @@ def main():
 
     trainer = GRPOTrainer(
         model=model,
-        ref_model=None, # let TRL create a frozen copy internally
+        ref_model=None,  # let TRL create a frozen copy internally
         args=cfg,
         tokenizer=tok,
         train_dataset=ds
@@ -470,7 +465,7 @@ def _write_trainer_script():
 
 # ---------------- Gradio Interface ----------------
 with gr.Blocks(theme=gr.themes.Soft()) as demo:
-    gr.Markdown("## ✍🏾 wilson Handwritten OCR — with Feedback Loop, Memory & GRPO Export")
+    gr.Markdown("## ✍🏾 Wilson Handwritten OCR — with Feedback Loop, Memory & GRPO Export")
 
     model_choice = gr.Radio(choices=list(MODEL_PATHS.keys()),
                             value=list(MODEL_PATHS.keys())[0],
@@ -573,14 +568,12 @@ with gr.Blocks(theme=gr.themes.Soft()) as demo:
 
         csv_status = gr.Markdown()
 
-# ------------------- MODIFIED -------------------
         gr.Markdown("---")
         gr.Markdown("### ⬇️ Download Feedback Data")
         with gr.Row():
             download_csv_btn = gr.Button("⬇️ Download Feedback as CSV")
             download_csv_file = gr.File(label="CSV File")
         download_csv_btn.click(fn=get_csv_file, outputs=download_csv_file)
-# ------------------------------------------------
 
     with gr.Tab("🧪 GRPO / Dataset"):
         gr.Markdown("""
@@ -591,35 +584,21 @@ with gr.Blocks(theme=gr.themes.Soft()) as demo:
 ```bash
 pip install trl accelerate peft transformers datasets
 python train/grpo_train.py
-````
 
-Set `BASE_MODEL`/`OUTPUT_DIR` env vars if you like.
+Set BASE_MODEL/OUTPUT_DIR env vars if you like.
 """)
-grpo\_btn = gr.Button("📦 Export GRPO Preferences")
-grpo\_status = gr.Markdown()
-grpo\_btn.click(fn=export\_grpo\_preferences, outputs=[grpo\_status])
-
-```
-    write_script_btn = gr.Button("📝 Write grpo_train.py")
-    write_script_status = gr.Markdown()
-    write_script_btn.click(fn=lambda: f"✅ Trainer script written to `{_write_trainer_script()}`", outputs=[write_script_status])
-```
-
-# \------------------- NEW -------------------
-
-```
-    gr.Markdown("---")
-    gr.Markdown("### ⬇️ Download GRPO Dataset")
-    with gr.Row():
-        download_grpo_btn = gr.Button("⬇️ Download GRPO Data (grpo_prefs.jsonl)")
-        download_grpo_file = gr.File(label="GRPO Dataset File")
-    download_grpo_btn.click(fn=get_grpo_file, outputs=download_grpo_file)
-```
-
-# \-------------------------------------------
-
-if **name** == "**main**":
-demo.queue(max\_size=50).launch(share=True)
-
-```
-```
+grpo_btn = gr.Button("📦 Export GRPO Preferences")
+grpo_status = gr.Markdown()
+grpo_btn.click(fn=export_grpo_preferences, outputs=[grpo_status])
+write_script_btn = gr.Button("📝 Write grpo_train.py")
+write_script_status = gr.Markdown()
+write_script_btn.click(fn=lambda: f"✅ Trainer script written to {_write_trainer_script()}", outputs=[write_script_status])
+gr.Markdown("---")
+gr.Markdown("### ⬇️ Download GRPO Dataset")
+with gr.Row():
+download_grpo_btn = gr.Button("⬇️ Download GRPO Data (grpo_prefs.jsonl)")
+download_grpo_file = gr.File(label="GRPO Dataset File")
+download_grpo_btn.click(fn=get_grpo_file, outputs=[download_grpo_file])
+
+if name == "main":
+demo.queue(max_size=50).launch(share=True)