Upload 18 files

app.py

@@ -158,16 +158,31 @@ _SELFTEST = {"done": False, "result": ""}
 
 
 def _export_dataset():
+    n = finetune_data.count()
+    if n == 0:
+        return ("⚠️ **0 training pairs captured yet.** Pairs are saved only when "
+                "the **Signals → AI summary** finishes with a model loaded. Steps: "
+                "1) Model tab — wait for the Translator sub-agent to show ✅; "
+                "2) Signals — Run analysis, pick a ticker, click **AI summary**, "
+                "let it finish; repeat a few times; 3) come back and Export.",
+                gr.update(visible=False))
     path = finetune_data.export()
     if not path:
-        return ("No training pairs captured yet. Run the Signals **AI summary** on "
-                "a few tickers first — each run is saved automatically.",
+        return ("⚠️ Export failed to write the file (storage error). Try again.",
                 gr.update(visible=False))
-    n = finetune_data.count()
-    msg = (f"✅ Exported dataset ready below. Captured {n} pair(s). "
-           f"Download it, then follow `finetune/FINETUNE_GUIDE.md` to LoRA-tune "
-           f"Qwen3-1.7B and publish it.")
-    return msg, gr.update(value=path, visible=True)
+    # Copy to a folder under the app's working dir, which Gradio serves
+    # reliably (the /data bucket and /tmp are not in Gradio's allowed paths).
+    import shutil
+    served_dir = os.path.join(os.getcwd(), "exports")
+    os.makedirs(served_dir, exist_ok=True)
+    served = os.path.join(served_dir, os.path.basename(path))
+    try:
+        shutil.copy(path, served)
+    except OSError:
+        served = path
+    msg = (f"✅ Exported **{n}** captured pair(s). Download below, then follow "
+           f"`finetune/FINETUNE_GUIDE.md` to LoRA-tune Qwen3-1.7B and publish it.")
+    return msg, gr.update(value=served, visible=True)
 
 
 def _run_selftest():
@@ -348,12 +363,11 @@ with gr.Blocks(title="Chan Compass · US", **_style_kw) as demo:
                     "few hundred, export the JSONL, then follow `FINETUNE_GUIDE.md` "
                     "to LoRA-tune Qwen3-1.7B and publish it.")
         ft_status = gr.Markdown(finetune_data.status_line())
-        with gr.Row():
-            ft_refresh = gr.Button("↻ Count pairs")
-            ft_export = gr.Button("⬇ Export dataset (JSONL)", variant="primary")
+        ft_export = gr.Button("⬇ Export dataset (JSONL)", variant="primary")
         ft_out = gr.Markdown()
         ft_file = gr.File(label="Download training data", visible=False)
-        ft_refresh.click(lambda: finetune_data.status_line(), None, ft_status)
+        ft_timer = gr.Timer(3.0)
+        ft_timer.tick(lambda: finetune_data.status_line(), None, ft_status)
         ft_export.click(_export_dataset, None, [ft_out, ft_file])
 
     gr.Markdown("Chan Compass · educational tool, not investment advice · "
@@ -403,7 +417,10 @@ if os.environ.get("AUTO_LOAD_MODEL", "1") == "1":
     threading.Thread(target=_auto_load_model, daemon=True).start()
 
 if __name__ == "__main__":
+    import paths as _p
+    _allowed = [os.path.join(os.getcwd(), "exports"), _p.DATASET_DIR]
+    os.makedirs(_allowed[0], exist_ok=True)
     if _GR_MAJOR >= 6:
-        demo.launch(theme=theme, css=S2_CSS)
+        demo.launch(theme=theme, css=S2_CSS, allowed_paths=_allowed)
     else:
-        demo.launch()
+        demo.launch(allowed_paths=_allowed)

finetune_data.py

@@ -33,8 +33,9 @@ INSTRUCTION = ("You are an equity analyst. Based only on this factual read of a
 
 
 def _clean(text: str) -> str:
-    # strip the UI prefix and any stray think tags
+    # strip stray think tags and any "AI narrative ..." UI prefix
     text = re.sub(r"<think>.*?</think>", "", text, flags=re.S)
+    text = re.sub(r"^🤖\s*\*\*AI narrative[^\n]*\*\*\s*", "", text)
     text = text.replace("🤖 **AI narrative (Translator sub-agent · Qwen3-1.7B):**", "")
     return text.strip()