README.md

---
license: mit
datasets:
- Recompense/amazon-appliances-lite-data
language:
- en
tags:
- finance


---
# Amazon Appliances Product Price Predictor

A Bi-LSTM model trained to predict e-commerce product prices from textual descriptions.

---

## Model Details

- **Model type:** Bi-directional LSTM (Keras)  
- **Task:** Regression (price prediction)  
- **Input:** Product description (text)  
- **Output:** Predicted price (USD)  

---

## Intended Use

This model is designed to provide quick, approximate pricing for small-to-medium sized e-commerce catalogs where descriptions follow a consistent style (e.g., electronics or appliances). It **should not** be used:

- For precise financial appraisal or high-stakes bidding.  
- On descriptions with highly technical jargon the model wasn’t trained on.  

---

## Limitations

- **Domain sensitivity:** Trained on the `Recompense/amazon-appliances-lite-data` dataset—performance may degrade on other product categories.  
- **Short descriptions:** Very long or unstructured text may reduce accuracy.  
- **Price range:** Only learns the range present in the training data (~\$10–\$500).  

---

## Training

- **Dataset:** `Recompense/amazon-appliances-lite-data`  
- **Preprocessing:**  
  - Text vectorization (max 10 000 tokens)  
  - Embedding dimension: 128  
- **Architecture:**  
  1. Embedding → Bi-LSTM(64) → Bi-LSTM(64) → Dense(1)  
- **Optimizer:** Adam, learning rate 1e-3  
- **Epochs:** 20, batch size 32  

---

## Evaluation

- **Metric:** Root Mean Squared Logarithmic Error (RMSLE)  
- **Formula (display mode):**

$$
RMSLE = \sqrt{ \frac{1}{n} \sum_{i=1}^{n} \bigl(\log(1 + \hat{y}_i) - \log(1 + y_i)\bigr)^2 }
$$

- **Test RMSLE:** 0.51 on held-out validation set hit 90.4% on validation set with margin of $40 

---

## Files

- **`model_weights.h5`** – Trained Keras weights  
- **`model_config.json`** – Model architecture config  
- **`vectorizer_config.json`** – Text vectorization config  

---

## Usage

Below is an end-to-end example showing how to load the model from the Hugging Face Hub, set your preferred Keras backend, and run inference using the helper function:

```python
# 1) Install dependencies (if needed)
#    pip install tensorflow jax keras huggingface_hub

# 2) Choose your backend: "jax", "torch", or "tensorflow"
import os
os.environ["KERAS_BACKEND"] = "jax"  # or "torch", or "tensorflow"

# 3) Load Keras and the model from the Hub
from keras.saving import load_model

model = load_model("hf://Recompense/product-pricer-bilstm")

# 4) Define your inference function
import tensorflow as tf

def bilstm_pricer(item_text: str) -> int:
    """
    Predict the price of a product given its description.
    
    Args:
        item_text (str): The full prompt text, including any prefix. 
                         Only the description (after the first blank line) is used.
    
    Returns:
        int: The rounded, non-negative predicted price in USD.
    """
    # Extract just the product description (assuming a prefix question)
    try:
        description = item_text.split('\n\n', 1)[1]
    except IndexError:
        description = item_text
    
    # Vectorize and batch the text
    text_tensor = tf.convert_to_tensor([description])
    
    # Model prediction
    pred = model.predict(text_tensor, verbose=0)[0][0]
    
    # Post-process: clamp and round
    pred = max(0.0, pred)
    return round(pred)

# 5) Example inference
prompt = (
    "What is a fair price for the following appliance?\n\n"
    "Stainless steel 12-cup programmable coffee maker with auto-shutoff"
)

predicted_price = bilstm_pricer(prompt)
print(f"Predicted price: ${predicted_price}")
```