NLP Sentiment & Text Generator
Budget: ₹600 – ₹2,500 INR
I’m sitting on a collection of structured, largely numerical features that represent language content (think embeddings, ratings, coded survey answers, or other vectorised traces of text). My goal is to turn those numbers back into meaningful language insight and output. Concretely, I need two tightly-linked models:
1. a sentiment classifier that can tag each record as positive, negative, or neutral, and
2. a text generator able to produce short, coherent passages consistent with the underlying signals in the data.
You’ll take the raw vectors, handle any additional preprocessing, select suitable Natural Language Processing architectures—Transformers, LSTM hybrids, or anything you can justify—then train, validate, and fine-tune until performance is solid.
Helpful toolkits are totally up to you; I’m comfortable with Python, PyTorch or TensorFlow, Hugging Face Transformers, and scikit-learn if they make the pipeline smoother. Deliverables I expect:
• Jupyter notebooks or .py scripts with clear, reproducible code
• Trained model weights and tokenizer files
• A concise README explaining environment setup and how to run inference
• Brief evaluation report showing at least 0.85 macro-F1 on sentiment and competitive perplexity/BLEU for generation
If you have ideas for pushing metrics even higher or for deploying the models behind a lightweight API, let’s discuss—I’m flexible as long as the core sentiment analysis and text generation goals are met.
1. a sentiment classifier that can tag each record as positive, negative, or neutral, and
2. a text generator able to produce short, coherent passages consistent with the underlying signals in the data.
You’ll take the raw vectors, handle any additional preprocessing, select suitable Natural Language Processing architectures—Transformers, LSTM hybrids, or anything you can justify—then train, validate, and fine-tune until performance is solid.
Helpful toolkits are totally up to you; I’m comfortable with Python, PyTorch or TensorFlow, Hugging Face Transformers, and scikit-learn if they make the pipeline smoother. Deliverables I expect:
• Jupyter notebooks or .py scripts with clear, reproducible code
• Trained model weights and tokenizer files
• A concise README explaining environment setup and how to run inference
• Brief evaluation report showing at least 0.85 macro-F1 on sentiment and competitive perplexity/BLEU for generation
If you have ideas for pushing metrics even higher or for deploying the models behind a lightweight API, let’s discuss—I’m flexible as long as the core sentiment analysis and text generation goals are met.