Update app.py

app.py

@@ -1,7 +1,7 @@
-# app.py — Gradio UI
-# - Train tab: streams live logs while running train.py
-# - After training, artifacts are uploaded to Files tab (handled by train.py)
-# - Use tab: pick trained adapter, load it (no fallback), generate strict JSON
+# app.py — Gradio UI with:
+#  - Train tab: live log streaming while running train.py
+#  - Use tab: pick a trained adapter, load (no fallback), generate strict JSON
+#  - Downloads: provides artifacts/<label>.zip for the loaded adapter
 
 import os, re, json, time, sys, subprocess
 from typing import Optional, Tuple
@@ -46,13 +46,15 @@ def adapter_exists(label: str) -> bool:
     has_info= "MODEL_INFO.json" in files
     return has_cfg and has_wts and has_info
 
+
 # --------------------- global loaded model ---------------------
 TOK = None
 MODEL = None
 ACTIVE_LABEL: Optional[str] = None
 ACTIVE_BASE = BASE_MODEL_DEFAULT
 
-# --------------------- training (streamed logs) ---------------------
+
+# --------------------- training (live logs) ---------------------
 def train_model_live(dataset_id, base_model, label, epochs):
     if not label:
         label = time.strftime("run_%Y%m%d_%H%M%S")
@@ -64,8 +66,10 @@ def train_model_live(dataset_id, base_model, label, epochs):
         "--label", label.strip(),
         "--epochs", str(int(epochs)),
     ]
+    # Show command first
     yield f"$ {' '.join(cmd)}\n\n"
 
+    # Stream stdout line-by-line
     process = subprocess.Popen(
         cmd,
         stdout=subprocess.PIPE,
@@ -81,10 +85,10 @@ def train_model_live(dataset_id, base_model, label, epochs):
             break
         if line:
             buffer += line
-            # stream progressively
-            yield buffer[-6000:]  # keep the tail to avoid giant textbox
+            yield buffer[-8000:]  # keep tail visible
     rc = process.wait()
-    yield buffer[-6000:] + f"\n\n[exit code: {rc}]"
+    yield buffer[-8000:] + f"\n\n[exit code: {rc}]"
+
 
 # --------------------- loading + generation ---------------------
 def load_selected_model(label: str) -> str:
@@ -99,7 +103,7 @@ def load_selected_model(label: str) -> str:
         meta = json.load(open(info_path, "r", encoding="utf-8"))
         base = meta.get("base_model", BASE_MODEL_DEFAULT)
     except Exception:
-        base = BASE_MODEL_DEFAULT
+        meta, base = {}, BASE_MODEL_DEFAULT
 
     TOK = AutoTokenizer.from_pretrained(base)
     base_model = AutoModelForSeq2SeqLM.from_pretrained(base)
@@ -107,7 +111,8 @@ def load_selected_model(label: str) -> str:
     MODEL.eval()
     ACTIVE_LABEL = label
     ACTIVE_BASE  = base
-    return f"✅ Loaded: {label} (base={base})"
+    ts = time.strftime("%Y-%m-%d %H:%M:%S", time.localtime(meta.get("saved_at", 0))) if meta else "unknown"
+    return f"✅ Loaded: {label} (base={base}, saved={ts})"
 
 PROMPT_HEADER = (
     "You are a quiz question generator. "
@@ -129,7 +134,8 @@ def try_parse_json(js: str) -> Optional[dict]:
     try:
         return json.loads(js)
     except Exception:
-        js2 = re.sub(r",\s*([}\]])", r"\\1", js)
+        # minor cleanup for trailing commas
+        js2 = re.sub(r",\s*([}\]])", r"\1", js)
         try:
             return json.loads(js2)
         except Exception:
@@ -144,7 +150,7 @@ def generate(age_band, genre, difficulty):
     out = MODEL.generate(
         **inputs,
         max_new_tokens=160,
-        do_sample=False,        # deterministic
+        do_sample=False,        # deterministic, cleaner JSON
         num_beams=4,
         early_stopping=True,
         no_repeat_ngram_size=3,
@@ -177,13 +183,15 @@ def do_load(label):
     can_gen = TOK is not None and MODEL is not None and ACTIVE_LABEL == label
     return status, (zip_path if os.path.isfile(zip_path) else None), gr.update(interactive=can_gen)
 
+
 # --------------------- UI ---------------------
 with gr.Blocks() as demo:
-    gr.Markdown("## Quiz AI — Train ➜ Save ➜ Download ➜ Use")
+    gr.Markdown("## Quiz AI — Train ➜ Save ➜ Download ➜ Use (No tokens)")
 
     with gr.Tab("Train"):
         with gr.Row():
-            dataset = gr.Textbox(value=os.environ.get("DATASET_ID", "Percy3822/quiz_ai_dataset"), label="Dataset ID (HF)", scale=2)
+            dataset = gr.Textbox(value=os.environ.get("DATASET_ID", "Percy3822/quiz_ai_dataset"),
+                                 label="Dataset ID (HF) — ignored if data/train.jsonl exists", scale=2)
             base    = gr.Textbox(value=BASE_MODEL_DEFAULT, label="Base model", scale=2)
         with gr.Row():
             label   = gr.Textbox(placeholder="e.g., quiz_v1 (leave blank for timestamp)", label="Adapter label", scale=2)