| --- |
| library_name: transformers |
| tags: |
| - finance |
| license: mit |
| datasets: |
| - Recompense/amazon-appliances-lite-data |
| language: |
| - en |
| base_model: |
| - meta-llama/Llama-3.1-8B-Instruct |
| --- |
| |
| # Model Card for Model ID |
|
|
| Predicts Prices based on product description. |
|
|
|
|
|
|
| ## Model Details |
|
|
| ### Model Description |
|
|
| This model predicts prices of amazon aplliances data based on a product description |
|
|
|
|
| - **Developed by:** https://huggingface.co/Recompense |
| - **Model type:** Transformer (causal, autoregressive) |
| - **Language(s) (NLP):** English |
| - **License:** MIT |
| - **Finetuned from model:** meta-llama/Llama-3.1-8B-Instruct |
|
|
| ### Model Sources |
|
|
| - **Repository:** [https://huggingface.co/Recompense/Midas-pricer/] |
|
|
| ## Uses |
|
|
| - Primary use case: Generating estimated retail prices for household appliances from textual descriptions. |
|
|
| - Example applications: |
|
|
| Assisting e-commerce teams in setting competitive price points |
|
|
| Supporting market analysis dashboards with on-the-fly price estimates |
|
|
| - Not intended for: Financial advice or investment decisions |
|
|
|
|
| ### Out-of-Scope Use |
|
|
| - Attempting to predict prices outside the appliances domain (e.g., electronics, furniture, vehicles) will likely yield unreliable results. |
|
|
| - Using this model for any price-sensitive or regulatory decisions without human oversight is discouraged. |
|
|
| ## Bias, Risks, and Limitations |
|
|
| - Data biases: The training dataset is drawn exclusively from Amazon appliance listings. Price distributions are skewed toward mid-range consumer electronics; extreme low or high‐end appliances are underrepresented. |
|
|
| - Input sensitivity: Minor changes in phrasing or additional noisy tokens can shift predictions noticeably. |
|
|
| - Generalization: The model does not understand supply chain disruptions, seasonality, or promotions—it only captures patterns seen in historical listing data. |
|
|
|
|
| ### Recommendations |
|
|
| - Always validate model outputs against a small set of ground-truth prices before production deployment. |
|
|
| - Use this model as an assistant, not an oracle: incorporate downstream business rules or domain expertise. |
|
|
| - Regularly retrain or fine-tune on updated listing data to capture shifting market trends. |
|
|
| ## How to Get Started with the Model |
|
|
| ```python |
| from transformers import AutoModelForCausalLM, AutoTokenizer |
| import torch |
| |
| # Load tokenizer and model |
| tokenizer = AutoTokenizer.from_pretrained("Recompense/Midas-pricer") |
| model = AutoModelForCausalLM.from_pretrained("Recompense/Midas-pricer", torch_dtype=torch.bfloat16) |
| |
| # Prepare prompt |
| product_desc = "How much does this cost to the nearest dollar?\n\nSamsung 7kg top-load washing machine with digital inverter motor" |
| prompt = f"{product_desc}\n\nPrice is $" |
| |
| # Tokenize and generate |
| inputs = tokenizer(prompt, return_tensors="pt") |
| attention_mask = torch.ones(inputs.shape, device="cuda") |
| generated = model.generate(inputs, attention_mask=attention_mask, max_new_tokens=3, num_return_sequences=1) |
| price_text = tokenizer.decode(generated[0], skip_special_tokens=True) |
| |
| print(f"Estimated price: ${price_text}") |
| ``` |
|
|
|
|
| ## Training Details |
|
|
| ### Training Data |
|
|
| - Dataset: Recompense/amazon-appliances-lite-data |
|
|
| - Train/validation/test split: 80/10/10 |
|
|
| ### Training Procedure |
|
|
| #### Training Hyperparameters |
|
|
| - Fine-tuning framework: PyTorch + Hugging Face Accelerate |
|
|
| - Precision: bf16 mixed precision |
|
|
| - Batch size: 1 sequence |
|
|
| - Learning rate: 1e-5 with linear warmup (10% of total steps) |
|
|
| - Optimizer: AdamW |
|
|
| ## Evaluation |
|
|
|
|
| ### Testing Data, Factors & Metrics |
| - Test set: Held-out 10% of listings (≈5 000 examples) |
|
|
| - Metric: Root Mean Squared Logarithmic Error (RMSLE) |
|
|
| - Hit@$40: Percentage of predictions within ±$40 of true price |
|
|
| | Metric | Value | |
| | -------- | ------ | |
| | RMSLE | 0.61 | |
| | Hit@\$40 | 85.2 % | |
|
|
|
|
| #### Summary |
| The model achieves an RMSLE of 0.61, indicating good alignment between predicted and actual prices on a log scale, |
| and correctly estimates within $40 in over 85% of test cases. |
| This performance is competitive for rapid prototyping in price-sensitive applications. |
|
|
| ## Environmental Impact |
|
|
| - Approximate compute emissions for fine-tuning (using ML CO₂ impact calculator): |
|
|
| - Hardware: Tesla T4 |
|
|
| - Duration: 2 hours(0.06 epoch) |
|
|
| - Cloud provider: Google Cloud, region US-Central |
|
|
| - Estimated CO₂ emitted: 6 kg CO₂e |
|
|
| ## Technical Specifications |
| ## Model Architecture |
| - Base model: Llama-3.1-8B (8 billion parameters) |
|
|
| - Objective: Autoregressive language modeling with instruction tuning |
|
|
| ## Compute Infrastructure |
| - Hardware: 4× Tesla T4 GPUs |
|
|
| - Software: |
|
|
| PyTorch 2.x |
|
|
| transformers 5.x |
|
|
| accelerate 1.x |
|
|
| bitsandbytes (for 8-bit quantization optional inference) |
|
|
|
|
|
|
| ## Glossary |
|
|
| - RMSLE (Root Mean Squared Logarithmic Error): |
| Measures the square root of the average squared difference between log-transformed predictions and targets. Less sensitive to large absolute errors. |
|
|
| - Hit@$40: |
| Fraction of predictions whose absolute error is ≤ $40. |
|
|
|
|
| ## Model Card Authors |
|
|
| Damola Jimoh(Recompense) |
