sentimentizer

Lightweight PyTorch models for sentiment analysis. Small models can be pretty effective for classification tasks at a much smaller cost to deploy — all models were trained on a single GPU in minutes, and inference requires less than 1GB of memory.

Beta release — API is subject to change.

Install

# Install local-only version (no Ray dependency)
uv add sentimentizer

# Install with distributed training, tuning, and serving features
uv add "sentimentizer[ray]"

Quick Start

Run a pre-trained model locally:

from sentimentizer.predictor import SentimentPredictor

# Load the model
predictor = SentimentPredictor(model_name="encoder")

# Predict sentiment (returns label, score, all-class scores, token count, and model type)
result = predictor.predict("amazing restaurant!")
# >> {"positive": 0.848753035068512,
#     "label": "positive",
#     "score": 0.848753035068512,
#     "scores": {"negative": 0.05746544152498245,
#                "neutral": 0.09378153085708618,
#                "positive": 0.848753035068512},
#     "token_count": 2,
#     "model": "encoder"}

# Batch prediction
results = predictor.predict_batch(["Great food!", "Terrible service."])
# >> [{'positive': 0.848753035068512,
#      'label': 'positive',
#      'score': 0.848753035068512,
#      'scores': {'negative': 0.05746544152498245,
#                 'neutral': 0.09378153085708618,
#                 'positive': 0.848753035068512},
#      'token_count': 2,
#      'model': 'encoder'},
#     {'negative': 0.9586858749389648,
#      'label': 'negative',
#      'score': 0.9586858749389648,
#      'scores': {'negative': 0.9586858749389648,
#                 'neutral': 0.03224344924092293,
#                 'positive': 0.0090706842020154},
#      'token_count': 2,
#      'model': 'encoder'}]

Models output 3-class probabilities (negative, neutral, positive) that sum to 1.0 per sample.

Intent Router API

The intent router classifies reviews as dietary, service, or general. It uses a SentenceTransformer backbone and LogisticRegression head; the former SetFit implementation has been replaced. Install serving and router support with uv add "sentimentizer[ray,router]", then start the API with sentimentizer serve. The router loads on the first classification request.

For example, send {"text": "They were so careful with my celiac needs"} to /v1/router/predict. The JSON response has this shape (scores and latency are illustrative):

{
  "prediction": {
    "label": "dietary",
    "score": 0.95,
    "token_count": 8
  },
  "latency_s": 0.0031
}

Batch requests accept {"texts": ["Great service", "Delicious pizza"]}; each entry in results uses the same {"prediction": {...}} envelope. score represents the predicted category's probability, token_count counts regex tokens, and latency_s measures server-side handler time in seconds. HTTP envelopes live in sentimentizer/serve/models.py; shared task and classification schemas live in sentimentizer/router/schemas.py.

See the complete JSON examples and validation rules and the router guide. Responses validate finite scores in [0, 1], nonnegative counts/timing, and complete batches. Confidence columns follow the classifier's saved class order.

The router uses task configuration and artifact-defined labels while retaining food-review defaults. To train another task, supply an external YAML config, its labeled data, and optional augmentation seeds. Each trained artifact saves its task definition; changing YAML alone does not retrain existing weights. See reuse with external configurations.

Models

Four architectures are available:

All models output 3-class logits (B, 3) mapped to: negative (0), neutral (1), positive (2).

Documentation

Detailed guides and implementation details are available in the specialized documentation files:

• 🚀 Model Serving Guide: Ray Serve application deployment, FastAPI endpoints (sentiment/routing/image generation), and the Go CLI client.
• 🏋️ Model Training & Checkpointing Guide: Yelp datasets, single-node/distributed commands, training arguments, sleep prevention, and checkpoint resuming.
• ⚙️ Model Configuration Reference: Configuration dataclasses (RNNConfig, EncoderConfig, etc.), parameter defaults, and consistency checks.
• 🎛️ Hyperparameter Tuning Guide: Optuna searches, LangGraph iterative agent tuning (via Ollama GLM 5.1), and validation/retries.
• 🔗 Hugging Face Hub Integration: Pre-trained weights synchronization, explicit pull/push, and auto-generated model cards.
• 📈 Metrics and Monitoring Pipeline: Exporter details, Grafana dashboards, Prometheus scrape targets, NaN handling, and real-time intra-epoch batch metrics.
• 🧭 Intent Review Router: Dietary/service/general classification, Python and JSON response contracts, Ollama GLM 5.1 augmentation, training, and evaluation.
• 🛠️ Troubleshooting Guide: Solutions for common issues like majority-class collapses, vocabulary matches, or scheduling.

Development

This project uses uv for dependency management.

# Install uv (if not already installed)
curl -LsSf https://astral.sh/uv/install.sh | sh

# Install dependencies (CPU-only PyTorch, no Ray)
uv sync

# Install with Ray distributed features
uv sync --extra ray

# Install dev and test suites
uv sync --extra dev --extra ray

# Full development install
uv sync --extra dev --extra ray --extra router --extra onnx

Local CUDA / GPU development

The locked packages resolve CPU-only PyTorch. To install CUDA-enabled PyTorch locally:

uv sync --no-sources-package torch

Note: This ignores CPU overrides in pyproject.toml and pulls PyTorch from PyPI with CUDA/NVIDIA libraries. Avoid committing changes to uv.lock.

Testing

Ensure local CI tests pass prior to submitting changes:

# Run all tests
uv run pytest tests/

# Run only Ray Train tests
uv run pytest tests/ -k "Ray"

# Run with coverage report
uv run pytest tests/ --cov=sentimentizer --cov-report=term-missing

# Verbose per-test output (or: make test-verbose)
uv run pytest tests/ -v

Project Structure

sentimentizer/
├── __init__.py          # Logging and timing utilities
├── compat.py            # Transformers/setfit compatibility shims
├── config.py            # Configuration dataclasses and constants
├── data_source.py       # Unified DataSource protocol (pandas/Ray)
├── device.py            # Device detection (cuda/mps/cpu)
├── env.py               # Environment setup (NVIDIA LD_LIBRARY_PATH)
├── extractor.py          # Ray Data extraction from zip/tar archives
├── exporter.py           # Standalone Prometheus metrics exporter
├── export_onnx.py        # ONNX export, quantization, validation
├── hf.py                # Hugging Face Hub push/pull + model card generation
├── hf_dataset.py        # Dataset wrapper and collation for HF transformers
├── hf_tokenizer.py      # Tokenizer wrapper for HF transformers
├── loader.py             # Data loading utilities
├── losses.py             # FocalCrossEntropyLoss for 3-class training
├── metrics.py            # 3-class classification metrics (per-class P/R/F1, balanced accuracy, MCC)
├── metrics_publisher.py   # Epoch metrics publishing (Prometheus + JSON) + intra-epoch batch snapshots
├── predictor.py           # SentimentPredictor (model loading, inference)
├── safety.py              # Prompt-injection patterns for web search results
├── serve/                 # Ray Serve deployment: FastAPI + @serve.ingress, /v1/ prefix
│   ├── app.py             # FastAPI route handlers and deployment class
│   ├── base.py            # ServiceMetrics (request/latency tracking), _DummyServe fallback
│   ├── config.py           # Serve deployment configuration (YAML/env var loading, incl. cors_origins)
│   └── models.py          # Pydantic request/response models for Swagger docs
├── tokenizer.py           # Text tokenizer with pre-trained support
├── trainer.py             # Training logic
├── tuner.py               # Ray Tune + Optuna hyperparameter search
├── data/                  # Training data (Yelp, GloVe)
├── agent/                 # LLM-guided tuning agent
│   ├── __init__.py       # Package exports
│   ├── config.yaml       # Agent + tuner configuration (YAML)
│   ├── loader.py         # YAML → dataclass config loader
│   ├── models.py         # Pydantic models (AnalysisResult, TuningDecision, etc.)
│   ├── agents.py         # Pydantic AI agents (GLM 5.1 via Ollama)
│   ├── prompts.py        # System prompts for analysis & strategy agents
│   ├── state.py          # LangGraph AgentState TypedDict
│   ├── nodes.py          # LangGraph node functions (analyze, decide, tune, evaluate)
│   ├── graph.py          # LangGraph StateGraph + run_agent_tuning() entry point
│   └── diagnose_model.py # TuningRun workflow (tune → train → validate → retry pipeline)
├── router/                # SentenceTransformer + LogisticRegression intent router
│   ├── __init__.py       # Package exports
│   ├── config.py         # RouterConfig (SetFitConfig alias), food defaults, AugmentConfig
│   ├── schemas.py        # RouterTaskSpec, RouterClassification, artifact metadata
│   ├── model.py          # RouterModel inference, save/load, legacy artifact migration
│   ├── seeds.py          # Golden example utterances per category
│   ├── augment.py        # GLM 5.1 augmentation via Ollama
│   ├── dataset.py        # JSONL dataset loader, train/test split
│   ├── train_router.py   # Contrastive backbone fine-tuning + classifier head fitting
│   └── evaluate.py       # Similarity heatmap, threshold calibration
└── models/
    ├── __init__.py
    ├── base.py            # BaseSentimentModel with predict() and predict_text()
    ├── hf_base.py         # Base class for Hugging Face transformer architectures
    ├── rnn.py            # Bidirectional LSTM (3-class output)
    ├── encoder.py         # Transformer encoder model (3-class output)
    ├── decoder.py         # Encoder-decoder transformer (3-class output)
    └── modernbert.py      # ModernBERT transformer classifier wrapper (3-class output)

License

MIT