Update app.py

app.py

@@ -1,5 +1,4 @@
 # app.py — HTR Space with Feedback Loop, Memory Post-Correction, and GRPO Export
-
 import os, time, json, hashlib, difflib, uuid, csv
 from datetime import datetime
 from collections import Counter, defaultdict
@@ -18,17 +17,17 @@ 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 3 removed to conserve memory.
 
 MAX_NEW_TOKENS_DEFAULT = 512
 device = "cuda" if torch.cuda.is_available() else "cpu"
@@ -161,7 +160,7 @@ 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", ""))
@@ -266,7 +265,7 @@ def _append_jsonl(path, obj):
 def _export_csv():
     # optional: CSV summary for spreadsheet views
     if not os.path.exists(FEEDBACK_PATH):
-        return CSV_EXPORT_PATH if os.path.exists(CSV_EXPORT_PATH) else None
+        return None
     rows = []
     with open(FEEDBACK_PATH, "r", encoding="utf-8") as f:
         for line in f:
@@ -285,15 +284,17 @@ 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:
-        return "Please provide the image again to link feedback.", 0
+        # 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.", 0
+        return "Nothing to save."
 
     image_hash = _hash_image(image)
     # best target = correction, else ground_truth, else prediction
@@ -319,7 +320,8 @@ def save_feedback(image: Image.Image, model_choice: str, prompt: str,
         "cer": float(cer_score) if cer_score is not None else None,
     }
     _append_jsonl(FEEDBACK_PATH, row)
-    return f"✅ Feedback saved (reward={reward}).", 1
+    return f"✅ Feedback saved (reward={reward})."
+# ------------------------------------------------
 
 def compile_memory_rules():
     _compile_rules_from_feedback(min_count=2, max_phrase_len=60)
@@ -357,11 +359,18 @@ def export_grpo_preferences():
                     count += 1
     return f"✅ Exported {count} GRPO preference pairs to {GRPO_EXPORT_PATH}."
 
-def export_csv():
-    p = _export_csv()
-    if p:
-        return f"✅ CSV exported: {p}"
-    return "No data to export."
+# ------------------- NEW -------------------
+def get_grpo_file():
+    if os.path.exists(GRPO_EXPORT_PATH):
+        return GRPO_EXPORT_PATH
+    return None
+
+def get_csv_file():
+    _export_csv()
+    if os.path.exists(CSV_EXPORT_PATH):
+        return CSV_EXPORT_PATH
+    return None
+# -------------------------------------------
 
 # ---------------- Evaluation Orchestration ----------------
 @spaces.GPU
@@ -438,7 +447,7 @@ 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
@@ -460,7 +469,7 @@ def _write_trainer_script():
 
 # ---------------- Gradio Interface ----------------
 with gr.Blocks(theme=gr.themes.Soft()) as demo:
-    gr.Markdown("## ✍🏾 wilson Handwritten text recognition with Feedback Loop")
+    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],
@@ -552,38 +561,61 @@ with gr.Blocks(theme=gr.themes.Soft()) as demo:
 
     with gr.Tab("✏️ Feedback & Memory"):
         gr.Markdown("""
-**Pipeline**  
-1) Save feedback (👍 / 👎) and add corrections.  
-2) Click **Build/Refresh Memory** to generate auto-fix rules from positive feedback.  
+**Pipeline**
+1) Save feedback (👍 / 👎) and add corrections.
+2) Click **Build/Refresh Memory** to generate auto-fix rules from positive feedback.
 3) Keep **Enable Memory Post-correction** checked on inference/eval tabs.
         """)
         build_mem_btn = gr.Button("🧠 Build/Refresh Memory from Feedback")
         mem_status = gr.Markdown()
         build_mem_btn.click(fn=compile_memory_rules, outputs=[mem_status])
 
-        csv_btn = gr.Button("📤 Export Feedback as CSV")
         csv_status = gr.Markdown()
-        csv_btn.click(fn=export_csv, outputs=[csv_status])
+
+# ------------------- 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("""
 **GRPO Fine-tuning** (run offline or in a training Space):
 - Click **Export GRPO Preferences** to produce `data/grpo_prefs.jsonl` of (prompt, chosen, rejected).
 - Click **Write Trainer Script** to create `train/grpo_train.py`.
-- Then run:  
+- Then run:
 ```bash
 pip install trl accelerate peft transformers datasets
 python train/grpo_train.py
-```
+````
+
 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])
+""")
+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)
+
 
-        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])
+# \-------------------------------------------
 
 if __name__ == "__main__":
     demo.queue(max_size=50).launch(share=True)