Upload README.md with huggingface_hub

README.md

@@ -1,55 +1,92 @@
 ---
-license: apache-2.0
-base_model: MB20261/Llama32-3B-Instruct-function-calling-1M
+license: llama3.2
+language:
+  - en
+library_name: transformers
 tags:
-- speculative-decoding-draft [dispatch-ai, mobile, quantized, gguf, phone-farm-tested]
+  - mobile
+  - on-device
+  - quantized
+  - gguf
+  - function-calling
+  - agent
+  - dispatchai
 pipeline_tag: text-generation
-language: [en]
 ---
+
 # Llama-3.2-1B-FunctionCall-mobile
-**Dispatch AI** — Built for mobile. Tested on real phones.
-## Category
-Function Calling — Q5 quality
-## Model
-Re-engineered from [MB20261/Llama32-3B-Instruct-function-calling-1M](https://huggingface.co/MB20261/Llama32-3B-Instruct-function-calling-1M).
-Size: 1926 MB. Q4_K_M GGUF for llama.cpp.
-## Usage
-```bash
-./llama-cli -m model.gguf -p "Hello" -n 100 -t 4 -c 512
-```
-🌐 [dispatchAI on HuggingFace](https://huggingface.co/dispatchAI)
 
+A 1B parameter Llama-3.2 model intended for function calling, optimized for mobile. 1926MB GGUF file that runs at 6.0 tokens/sec on CPU.
+
+## ⚠️ Partially Verified
+
+This model has been tested with mixed results:
+- **Chat format**: `chatml` (use `chat_format="chatml"` in llama-cpp-python)
+- **Test date**: June 2026
+
+### Chat Test Results
+
+| Prompt | Response | Correct? |
+|--------|----------|----------|
+| "What is the capital of France?" | "The capital of France is Paris." | ✅ |
+| "What is 2+2?" | "The answer is 4." | ✅ |
+| "Write a greeting." | "How are you doing today?" | ✅ |
+
+### Function Calling Test Results
+
+| Prompt | Response | Correct? |
+|--------|----------|----------|
+| "Set an alarm for 7am" | Refused, said it can only send messages | ❌ |
+| "Call my mom" | Generated unrelated JSON | ❌ |
+| "What's the weather in Dubai?" | `<functioncall> get_weather("Dubai")` | ✅ |
 
-## Speculative Decoding Draft Model
+### Honest Assessment
 
-This model is optimized for use as a **draft model** in speculative decoding setups.
+**Chat**: Works well — coherent, correct responses.
+**Function calling**: Inconsistent — this appears to be a base Llama-3.2-1B model, not actually fine-tuned for function calling. It can parse some intents with a strong system prompt but is unreliable. For production function calling, consider using a properly fine-tuned model or few-shot prompting with examples.
 
-### What is speculative decoding?
-Speculative decoding pairs a small, fast "draft" model with a larger "target" model.
-The draft model proposes tokens that the target model verifies in parallel, achieving
-2-3x speedup with zero quality loss.
+## Model Details
 
-### Why this model?
-- **Small and fast**: Sub-1B parameters = minimal draft overhead
-- **Mobile-optimized**: Already quantized and pruned for edge deployment
-- **Same family**: Pairs naturally with larger models of the same architecture
+| Attribute | Value |
+|-----------|-------|
+| **Base Model** | meta-llama/Llama-3.2-1B-Instruct |
+| **Parameters** | 1B |
+| **File Size** | 1926 MB |
+| **Format** | GGUF |
+| **Chat Format** | chatml |
+| **License** | Llama 3.2 License |
+
+## Usage
 
-### Usage with vLLM
 ```python
-from vllm import LLM, SamplingParams
+from llama_cpp import Llama
+
+llm = Llama(model_path="model.gguf", chat_format="chatml", n_ctx=512, n_threads=4)
+
+# Chat (works well)
+response = llm.create_chat_completion(
+    messages=[{"role": "user", "content": "What is the capital of France?"}],
+    max_tokens=50,
+)
 
-llm = LLM(
-    model="target-model-7b",
-    speculative_model="dispatchAI/Llama-3.2-1B-FunctionCall-mobile",
-    num_speculative_tokens=5,
+# Function calling (use with strong system prompt)
+response = llm.create_chat_completion(
+    messages=[
+        {"role": "system", "content": "You are a function-calling assistant. Parse the user request into a JSON function call. Available: set_alarm(time), call_contact(name), get_weather(location). Respond with ONLY the JSON."},
+        {"role": "user", "content": "What's the weather in Dubai?"}
+    ],
+    max_tokens=50,
+    temperature=0.1,
 )
 ```
 
-### Usage with transformers
-```python
-from transformers import AutoModelForCausalLM, AutoTokenizer
+## Limitations
 
-target = AutoModelForCausalLM.from_pretrained("target-model-7b")
-draft = AutoModelForCausalLM.from_pretrained("dispatchAI/Llama-3.2-1B-FunctionCall-mobile")
-# See transformers docs for assisted_generation
-```
+- Not actually fine-tuned for function calling — uses a base Llama-3.2-1B
+- Function calling is inconsistent (~33% success rate in testing)
+- 1926MB file is large for a 1B model (may be FP16 instead of quantized)
+- For reliable function calling, use few-shot prompting or a dedicated FC model
+
+## About dispatchAI
+
+[dispatchAI](https://huggingface.co/dispatchAI) — Small. Mobile. Free. UAE-built.