Upload README.md with huggingface_hub

README.md

@@ -1,50 +1,102 @@
 ---
-library_name: mlx
-license: other
-license_name: lfm1.0
-license_link: LICENSE
+license: apache-2.0
+base_model: LiquidAI/LFM2.5-1.2B-Instruct
+tags:
+- text-to-sql
+- sql
+- fine-tuned
+- mlx
+- lora
+datasets:
+- synthetic
 language:
 - en
-- ar
-- zh
-- fr
-- de
-- ja
-- ko
-- es
 pipeline_tag: text-generation
-tags:
-- liquid
-- lfm2.5
-- edge
-- mlx
-base_model: mlx-community/LFM2.5-1.2B-Instruct-4bit
 ---
 
-# hybridaione/LFM2.5-1.2B-Text2SQL
+# LFM2.5-1.2B-Text2SQL
+
+A fine-tuned version of [LiquidAI/LFM2.5-1.2B-Instruct](https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct) optimized for text-to-SQL generation.
+
+## Model Description
+
+This model was fine-tuned using LoRA on 2000 synthetic text-to-SQL examples generated via knowledge distillation from DeepSeek V3. The fine-tuning was performed using MLX on Apple Silicon.
 
-This model [hybridaione/LFM2.5-1.2B-Text2SQL](https://huggingface.co/hybridaione/LFM2.5-1.2B-Text2SQL) was
-converted to MLX format from [mlx-community/LFM2.5-1.2B-Instruct-4bit](https://huggingface.co/mlx-community/LFM2.5-1.2B-Instruct-4bit)
-using mlx-lm version **0.29.1**.
+## Performance
 
-## Use with mlx
+| Metric | Teacher (DeepSeek V3) | Base (LFM2.5 1.2B) | This Model |
+|--------|----------------------|-------------------|------------|
+| **Exact Match** | 60% | 48% | **66%** |
+| **LLM-as-Judge** | 90% | 75% | 87% |
+| **ROUGE-L** | 0.917 | 0.830 | **0.931** |
+| **BLEU** | 0.852 | 0.695 | **0.870** |
+| **Semantic Similarity** | 0.965 | 0.926 | **0.970** |
+
+The fine-tuned model **beats the teacher on 4 out of 5 metrics** despite being significantly smaller.
+
+## Training Details
+
+- **Base Model:** LiquidAI/LFM2.5-1.2B-Instruct
+- **Fine-tuning Method:** LoRA (rank 8)
+- **Training Data:** 2000 synthetic examples
+- **Epochs:** 2 (checkpoint 1800)
+- **Hardware:** Apple Silicon (MLX)
+
+## Usage
+
+### With vLLM
+
+```python
+from vllm import LLM, SamplingParams
 
-```bash
-pip install mlx-lm
+llm = LLM(model="hybridaione/LFM2.5-1.2B-Text2SQL")
+sampling_params = SamplingParams(temperature=0, max_tokens=512)
+
+prompt = """<|im_start|>system
+You are an expert SQL writer. Given a database schema and natural language question, write the precise SQL query that answers it. Output only the SQL query with no explanation.<|im_end|>
+<|im_start|>user
+Schema:
+CREATE TABLE users (id INTEGER PRIMARY KEY, name TEXT, email TEXT);
+
+Question: How many users are there?<|im_end|>
+<|im_start|>assistant
+"""
+
+output = llm.generate([prompt], sampling_params)
+print(output[0].outputs[0].text)
 ```
 
+### With Transformers
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+model = AutoModelForCausalLM.from_pretrained("hybridaione/LFM2.5-1.2B-Text2SQL")
+tokenizer = AutoTokenizer.from_pretrained("hybridaione/LFM2.5-1.2B-Text2SQL")
+```
+
+### With MLX (Apple Silicon)
+
 ```python
 from mlx_lm import load, generate
 
 model, tokenizer = load("hybridaione/LFM2.5-1.2B-Text2SQL")
+response = generate(model, tokenizer, prompt="...", max_tokens=512)
+```
 
-prompt = "hello"
+## Prompt Format
 
-if tokenizer.chat_template is not None:
-    messages = [{"role": "user", "content": prompt}]
-    prompt = tokenizer.apply_chat_template(
-        messages, add_generation_prompt=True
-    )
+```
+<|im_start|>system
+You are an expert SQL writer. Given a database schema and natural language question, write the precise SQL query that answers it. Output only the SQL query with no explanation.<|im_end|>
+<|im_start|>user
+Schema:
+{CREATE TABLE statements}
 
-response = generate(model, tokenizer, prompt=prompt, verbose=True)
+Question: {natural language question}<|im_end|>
+<|im_start|>assistant
 ```
+
+## License
+
+Apache 2.0