Update README with usage instructions

README.md

@@ -1,136 +1,71 @@
 ---
 license: other
 license_name: qwen
-license_link: https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct/blob/main/LICENSE
-base_model: Qwen/Qwen3-VL-4B-Instruct
+license_link: https://huggingface.co/Qwen/Qwen3-VL-4B/blob/main/LICENSE
 tags:
+  - qwen3_vl
+  - image-to-text
   - vision-language-model
   - finance
-  - ocr
+  - OCR
   - chart-understanding
   - financial-analysis
-  - qwen3-vl
-language:
-  - en
-pipeline_tag: image-text-to-text
-library_name: transformers
 ---
 
 # Amsi-fin: Financial Vision-Language Model
 
-A specialized Vision-Language Model fine-tuned for financial document understanding, OCR, chart analysis, and chain-of-thought reasoning.
+Fine-tuned Qwen3-VL-4B for financial document understanding, chart analysis, and financial reasoning.
 
-## Model Details
+## Quick Start
 
-| Property | Value |
-|----------|-------|
-| **Base Model** | Qwen3-VL-4B-Instruct |
-| **Parameters** | 4 Billion |
-| **Precision** | BF16 |
-| **Context Length** | 131,072 tokens |
-| **Training Stages** | 4 (Progressive Fine-tuning) |
+### MLX (Apple Silicon)
 
-## Capabilities
-
-- **Financial Document OCR**: Extract text from financial reports, statements, and documents
-- **Chart Understanding**: Analyze and interpret financial charts and graphs
-- **Chain-of-Thought Reasoning**: Step-by-step financial analysis and calculations
-- **Mathematical Reasoning**: Financial calculations and numerical analysis
-
-## Training Data
-
-The model was trained on a curated mix of financial datasets:
+```python
+from mlx_vlm import load, generate
 
-| Stage | Focus | Datasets |
-|-------|-------|----------|
-| A1 | Foundation | FinTrain (70%), FinTrain-Math (15%), OCR (10%), ChartQA (5%) |
-| A2 | Vision/OCR | MultiFinBen-OCR (50%), SecureFinAI-OCR (20%), ChartQA (20%), NuminaMath (10%) |
-| A3 | Reasoning | FinCoT (60%), FinTrain (30%), OCR (5%), ChartQA (5%) |
-| A4 | Consolidation | FinTrain (40%), OCR (20%), FinCoT (20%), ChartQA (10%), NuminaMath (10%) |
+# IMPORTANT: Use fix_mistral_regex=True
+model, processor = load('AITRADER/Amsi-fin', fix_mistral_regex=True)
 
-## Training Configuration
+# Vision task
+output = generate(
+    model, processor, 
+    image='chart.png',
+    prompt='<|vision_start|><|image_pad|><|vision_end|>Analyze this chart.',
+    max_tokens=500
+)
 
-```yaml
-per_device_batch_size: 4
-gradient_accumulation_steps: 2
-learning_rate: 6.0e-6 (final stage)
-max_seq_length: 1024
-precision: bf16
-optimizer: AdamW (fused)
-total_steps: 7000 (across all stages)
+# Text-only
+output = generate(
+    model, processor,
+    prompt='Calculate debt-to-equity ratio if debt=120M, equity=80M.',
+    max_tokens=200
+)
 ```
 
-## Usage
-
-### With Transformers
+### Transformers (CUDA/CPU)
 
 ```python
 from transformers import AutoProcessor, AutoModelForVision2Seq
 import torch
 
-model_name = "AITRADER/Amsi-fin"
-
-processor = AutoProcessor.from_pretrained(model_name, trust_remote_code=True)
+processor = AutoProcessor.from_pretrained('AITRADER/Amsi-fin', trust_remote_code=True)
 model = AutoModelForVision2Seq.from_pretrained(
-    model_name,
+    'AITRADER/Amsi-fin',
     torch_dtype=torch.bfloat16,
-    device_map="auto",
     trust_remote_code=True
 )
-
-# Example: Analyze a financial document
-from PIL import Image
-
-image = Image.open("financial_report.png")
-messages = [
-    {"role": "user", "content": [
-        {"type": "image"},
-        {"type": "text", "text": "Analyze this financial report and summarize the key metrics."}
-    ]}
-]
-
-inputs = processor(messages, images=[image], return_tensors="pt").to(model.device)
-outputs = model.generate(**inputs, max_new_tokens=512)
-response = processor.decode(outputs[0], skip_special_tokens=True)
-print(response)
 ```
 
-### Convert to MLX (Apple Silicon)
-
-```bash
-# Install mlx-lm
-pip install mlx-lm
-
-# Convert to MLX 8-bit quantized
-mlx_lm.convert --hf-path AITRADER/Amsi-fin -q --upload-repo AITRADER/Amsi-fin-MLX-8bit
-
-# Convert to MLX bf16
-mlx_lm.convert --hf-path AITRADER/Amsi-fin --upload-repo AITRADER/Amsi-fin-MLX-bf16
-```
-
-## Limitations
-
-- Optimized for English financial documents
-- Best performance on structured financial data (tables, charts, reports)
-- May require fine-tuning for specific financial domains
-
-## License
-
-This model is released under the same license as the base Qwen3-VL model.
-
-## Citation
+## Capabilities
 
-```bibtex
-@misc{amsi-fin-2025,
-  title={Amsi-fin: Financial Vision-Language Model},
-  author={AITRADER},
-  year={2025},
-  publisher={HuggingFace},
-  url={https://huggingface.co/AITRADER/Amsi-fin}
-}
-```
+- Financial Document OCR
+- Chart/Graph Understanding  
+- Financial Reasoning & Calculations
+- Table Extraction
 
-## Acknowledgments
+## Training Data
 
-- Base model: [Qwen3-VL](https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct)
-- Training datasets: FinTrain, FinCoT, MultiFinBen, ChartQA, NuminaMath
+- FinTrain (Salesforce)
+- MultiFinBen-EnglishOCR
+- ChartQA
+- FinCoT