| --- |
| language: |
| - en |
| license: mit |
| tags: |
| - modernbert |
| - int4 |
| - qat |
| - nli |
| - zero-shot-classification |
| - sentiment-analysis |
| - jerry |
| pipeline_tag: zero-shot-classification |
| --- |
| |
| # General Purpose Jerry (Universal NLU & Zero-Shot Engine — INT4 W4A16) |
|
|
| **General Purpose Jerry** is the fine-tuned 4-bit NLU & Zero-Shot foundation model in the **Jerry Model Family** hosted at [`Amogh1221/Jerry`](https://huggingface.co/Amogh1221/Jerry). |
|
|
| - **Model Size**: **91.1 MB** (vs 220 MB FP16 base model — **75% reduction**). |
| - **Quantization**: **W4A16 (Weight-Only INT4 with Group-Wise QAT, group_size=64)**. |
| - **Pre-Trained Base**: 150k steps on ~9.8B tokens. |
| - **Fine-Tuning**: Multi-Genre Natural Language Inference (MNLI). |
| - **Validation Accuracy**: **80.77%**. |
| |
| ## Built-In Capabilities |
| 1. **Zero-Shot Sentiment Analysis**: |
| - Classifies positive, negative, and neutral sentiments with 0 training samples. |
| 2. **Zero-Shot Intent & Topic Classification**: |
| - Compares user prompts against custom candidate labels. |
| 3. **General NLU & Transfer Learning**: |
| - Plug-and-play feature extractor for high-speed downstream NLP pipelines. |
| |
| ## Inference Example (Zero-Shot Sentiment) |
| ```python |
| import torch |
| from jerry_infer import load_general_jerry, predict_zero_shot_sentiment |
| |
| classifier = load_general_jerry("general_jerry_int4.pt") |
| review = "The battery life on this laptop easily lasts 14 hours and the screen is gorgeous." |
| result = predict_zero_shot_sentiment(classifier, review) |
| print(result) # {'label': 'Positive', 'confidence': 0.94} |
| ``` |
| |
