Upload folder using huggingface_hub

README.md

@@ -1,13 +1,35 @@
 ---
 title: Tiny Browser Planner
-emoji: 📈
-colorFrom: pink
-colorTo: blue
+emoji: 🤖
+colorFrom: blue
+colorTo: purple
 sdk: gradio
-sdk_version: 6.18.0
-python_version: '3.13'
+sdk_version: 4.36.1
+python_version: 3.12.0
 app_file: app.py
 pinned: false
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# Tiny Browser Planner
+
+A 1B model that plans browser actions — fine-tuned MiniCPM5-1B with LoRA.
+
+**The core finding:** A 1B model can explain the correct browser action before it can reliably choose it.
+
+- [Model on HF](https://huggingface.co/Georgefifth/tiny-browser-planner-reason)
+- [Dataset on HF](https://huggingface.co/datasets/Georgefifth/tiny-browser-planner-reason-dataset)
+
+## Experiments
+
+### v1 → v4
+Data scaling didn't help. 200 targeted hard examples outperformed 2000 generic ones.
+
+### Ablation: Action Space Paradox
+Adding `back` to the action space created a powerful tool — but the model over-generalized it, using `back` for everything.
+
+### Reason-First
+With just 40 reasoning examples (7.8s training), the model went from 4/12 → 10/12. The reasoning head eliminates heuristic shortcutting.
+
+## Usage
+
+Enter a task and browsing history. The model shows its reasoning, then picks the next action.

app.py

@@ -0,0 +1,86 @@
+import re, time, json, os, shutil, torch, gradio as gr
+import tempfile
+from transformers import AutoModelForCausalLM, AutoTokenizer
+from peft import PeftModel
+from huggingface_hub import snapshot_download
+
+BASE_ID = "openbmb/MiniCPM5-1B"
+ADAPTER_ID = "Georgefifth/tiny-browser-planner-reason"
+
+print("Loading model (this may take a minute)...")
+start = time.time()
+
+# Download adapter and create clean config
+adapter_dir = os.path.join(tempfile.gettempdir(), "adapter")
+from huggingface_hub import snapshot_download
+snapshot_download(repo_id="Georgefifth/tiny-browser-planner-reason", local_dir=adapter_dir)
+with open(os.path.join(adapter_dir, "adapter_config.json")) as f:
+    raw_cfg = json.load(f)
+
+KEEP = {"r","lora_alpha","lora_dropout","target_modules","bias","task_type","peft_type","inference_mode"}
+clean_cfg = {k: v for k, v in json.load(open(os.path.join(adapter_dir, "adapter_config.json"))).items() if k in {"r","lora_alpha","lora_dropout","target_modules","bias","task_type","peft_type","inference_mode"}}
+
+clean_dir = os.path.join(tempfile.gettempdir(), "clean_adapter")
+os.makedirs(clean_dir, exist_ok=True)
+import shutil
+for fname in os.listdir(adapter_dir):
+    src = os.path.join(adapter_dir, fname)
+    dst = os.path.join(clean_dir, fname)
+    if os.path.isfile(src):
+        if fname == "adapter_config.json":
+            with open(dst, "w") as f:
+                json.dump({"r":16,"lora_alpha":16,"lora_dropout":0,"target_modules":["q_proj","k_proj","v_proj","o_proj","gate_proj","up_proj","down_proj"],"bias":"none","task_type":"CAUSAL_LM","peft_type":"LORA","inference_mode":True}, f)
+        else:
+            shutil.copy2(src, dst)
+
+model = AutoModelForCausalLM.from_pretrained("openbmb/MiniCPM5-1B", torch_dtype=torch.float16, trust_remote_code=True)
+model = PeftModel.from_pretrained(model, clean_dir)
+tokenizer = AutoTokenizer.from_pretrained("openbmb/MiniCPM5-1B", trust_remote_code=True)
+if tokenizer.pad_token is None:
+    tokenizer.pad_token = tokenizer.eos_token
+print(f"Ready! ({time.time()-start:.0f}s)")
+
+ACTIONS = ["search", "open_page", "extract", "refine_search", "back", "finish"]
+
+def predict(task, history_text):
+    if not task or not task.strip():
+        return "Error: task is empty", ""
+    history = [l.strip() for l in history_text.strip().split("\n") if l.strip()]
+    hist_str = "\n".join(history)
+    msgs = [
+        {"role": "system", "content": "You are a browser planner. First reason about the situation, then output the next action."},
+        {"role": "user", "content": f"Task: {task}\n\nHistory:\n{hist_str}\n\nWhat is the next action?"},
+    ]
+    prompt = tokenizer.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
+    inputs = tokenizer(prompt, return_tensors="pt")
+    input_len = inputs["input_ids"].shape[1]
+    inputs.pop("token_type_ids", None)
+    outs = model.generate(**inputs, max_new_tokens=64, temperature=0.01, do_sample=False,
+                          pad_token_id=tokenizer.eos_token_id)
+    output = tokenizer.decode(outs[0][input_len:], skip_special_tokens=True).strip()
+    reason_m = re.search(r"Reason:\s*(.+?)(?:\n|$)", output)
+    action_m = re.search(r"Action:\s*(\S+)", output)
+    reason = reason_m.group(1).strip() if reason_m else "?"
+    action = action_m.group(1).strip().lower() if action_m else "?"
+    if action not in ["search", "open_page", "extract", "refine_search", "back", "finish"]:
+        action = f"{action} (unknown)"
+    return reason, action
+
+with gr.Blocks(title="Tiny Browser Planner", theme=gr.themes.Soft()) as demo:
+    gr.Markdown("""
+    # Tiny Browser Planner — Reason-First
+    MiniCPM5-1B + LoRA  |  Actions: `search`, `open_page`, `extract`, `refine_search`, `back`, `finish`
+    """)
+    with gr.Row():
+        with gr.Column(scale=2):
+            task = gr.Textbox(label="Task", placeholder="e.g. Find Apple stock price")
+            history = gr.Textbox(label="History (one action per line)", lines=5,
+                                 placeholder="[search] Search completed.\n[open_page] Page content here...")
+            btn = gr.Button("Predict", variant="primary")
+        with gr.Column(scale=1):
+            reason = gr.Textbox(label="Reason", lines=3, interactive=False)
+            action = gr.Textbox(label="Next Action", lines=1, interactive=False)
+    btn.click(fn=predict, inputs=[task, history], outputs=[reason, action])
+
+if __name__ == "__main__":
+    demo.launch()

requirements.txt

@@ -0,0 +1,6 @@
+transformers>=4.30.0
+torch>=2.0.0
+peft>=0.6.0
+accelerate>=0.24.0
+safetensors>=0.4.0
+huggingface_hub>=0.20.0,<0.25.0