Upload folder using huggingface_hub

README.md

@@ -66,14 +66,19 @@ Liquid AI's LFM2.5-1.2B-Instruct model fine-tuned with LoRA adapters using Huggi
 | Training examples | 3,200 |
 | Test examples | 800 |
 | Epochs | 3 |
-| Batch size | 4 |
+| Batch size | 1 (effective 4 with gradient accumulation steps = 4) |
 | Learning rate | 2e-4 |
-| Gradient checkpointing | Enabled |
+| Max sequence length | 256 |
+| Optimizer | adamw_torch (bitsandbytes 8-bit not supported on MPS) |
+| Weight dtype | bfloat16 |
+| Device | MPS (Apple Silicon) |
+| Gradient checkpointing | Enabled (use_reentrant=False) |
+| Max gradient norm | 0.3 |
 | LoRA rank | 8 |
 | LoRA alpha | 16 |
 | LoRA dropout | 0.1 |
 | Target modules | 8 (q_proj, k_proj, v_proj, out_proj, w1, w2, w3, in_proj) |
-| Training time | ~2–2.5 hours |
+| Training time | ~1–1.5 hours (per fine_tune.py; earlier docs listed ~2–2.5 hours before the v0.4.3 memory optimization) |
 
 ### Hardware
 
@@ -121,7 +126,7 @@ It is **not** intended for production spam filtering.
 
 ## Limitations
 
-- **Binary classification only** (spam/ham) — no multi-class or severity ranking (v1)
+- **Three-class classification** (SPAM / HAM / PHISHING) as of v0.4.0 — earlier versions were binary
 - Model is too large for free HuggingFace Spaces deployment
 - May misclassify legitimate marketing emails as spam
 - Trained on **English emails only** — not suitable for other languages

app.py

@@ -90,6 +90,31 @@ tokenizer = None
 
 adapter_exists = Path(ADAPTER_PATH).exists() and any(Path(ADAPTER_PATH).iterdir())
 
+# If local adapters are missing (e.g. running on HuggingFace Spaces),
+# download them from the HF model repo instead.
+HF_ADAPTER_REPO = "VoltageVagabond/spam-classifier-liquid"
+
+if not adapter_exists:
+    print(f"Local adapters not found. Downloading from {HF_ADAPTER_REPO}...")
+    try:
+        import os
+        from huggingface_hub import snapshot_download
+        snapshot_path = snapshot_download(
+            repo_id=HF_ADAPTER_REPO,
+            repo_type="model",
+            allow_patterns=["adapters_fast/adapter_config.json",
+                            "adapters_fast/adapter_model.safetensors",
+                            "adapters_fast/chat_template.jinja",
+                            "adapters_fast/tokenizer.json",
+                            "adapters_fast/tokenizer_config.json"],
+            token=os.environ.get("HF_TOKEN"),
+        )
+        ADAPTER_PATH = str(Path(snapshot_path) / "adapters_fast")
+        adapter_exists = True
+        print(f"Adapters downloaded to {ADAPTER_PATH}")
+    except Exception as e:
+        print(f"ERROR: Could not download adapters: {e}")
+
 if adapter_exists:
     print("Loading Liquid AI model and LoRA adapters...")
     base_model = AutoModelForCausalLM.from_pretrained(
@@ -143,6 +168,8 @@ def generate_response(messages, max_tokens=750):
             max_new_tokens=max_tokens,
             do_sample=True,
             temperature=0.1,
+            cache_implementation="quantized",
+            cache_config={"backend": "hqq", "nbits": 8},
         )
 
     # Decode only the NEW tokens (skip the input prompt)
@@ -370,6 +397,8 @@ TOPBAR_HTML = """
 
 with gr.Blocks(
     title="Liquid AI Spam Classifier",
+    theme=theme,
+    css=custom_css,
 ) as demo:
 
     gr.HTML(TOPBAR_HTML)
@@ -378,6 +407,7 @@ with gr.Blocks(
     chatbot = gr.Chatbot(
         label="Chat",
         height=450,
+        type='messages',
     )
 
     # Message input row
@@ -476,4 +506,4 @@ with gr.Blocks(
 # ---------------------------------------------------------------------------
 
 if __name__ == "__main__":
-    demo.launch(theme=theme, css=custom_css)
+    demo.launch()

requirements.txt

@@ -3,6 +3,7 @@ transformers>=5.0.0
 torch>=2.6.0
 accelerate>=1.0.0
 peft>=0.14.0
+hqq>=0.2.0
 gradio>=5.0
 numpy>=1.24.0
 pandas>=2.0.0