How xiaooolong/vev works
Vev is a self-hosted, Jev-compatible decision model: callers submit text or JSON state, optionally containing images, and receive probabilities over defined answers. The serving path accepts requests, renders state and questions, and scores answer tokens with a checkpoint; it can also be invoked offline through the model engine. The repository additionally contains data preparation, fine-tuning, and benchmark evaluation workflows. Major documented capabilities include image judgments, batched questions, and evaluation of accuracy and calibration. Some data and evaluation subsystems are represented by their enclosing directories because their individual implementations were not sampled.
Main entry points: model.py Checkpoint engine · build.py Dataset builder · data.py Training rows · serve.py HTTP service
Components 21
Inference and serving
- HTTP service
vev/serve.py - State parsing
vev/state.py - Answer readout
vev/readout.py - Checkpoint engine
vev/model.py
Training data
- Dataset converters
data/sources - Source helpers
data/sources/_common.py - Dataset builder
data/build.py - Record augmentation
data/augment.py - Dataset splitting
data/split.py - Build records
data
Model training
- Training rows
train/data.py - Fine-tuning
train/train.py
Evaluation
- Benchmark converters
evals/datasets/_common.py - Evaluation datasets
evals/datasets - Evaluation runner
evals/run.py - Evaluation client
evals/client.py - Metrics
evals/metrics.py
Other components
- API caller
- Model weights
- Offline caller
- Decision service
Connections 24
- API caller → HTTP service: sends request
README.md - HTTP service → State parsing: parses state
vev/serve.py - HTTP service → Checkpoint engine: scores request
vev/serve.py - Checkpoint engine → State parsing: renders state
vev/model.py - Checkpoint engine → Answer readout: uses readout
vev/model.py - Answer readout → State parsing: uses segments
vev/readout.py - Checkpoint engine → Model weights: loads checkpoint
- Offline caller → Checkpoint engine: runs inference
README.md - Dataset converters → Source helpers: uses helpers
data/sources/text_amazon_reviews_multi_en.py - Source helpers → Benchmark converters: imports helpers
data/sources/_common.py - Source helpers → Dataset builder: feeds build
- Dataset builder → Record augmentation: imports augmentation
data/build.py - Dataset builder → Dataset splitting: imports splitting
data/build.py - Dataset builder → Build records: writes records
data/build.py - Training rows → Build records: reads build data
train/data.py - Training rows → Record augmentation: applies augmentation
train/data.py - Training rows → Checkpoint engine: renders rows
train/data.py - Fine-tuning → Training rows: loads rows
train/train.py - Fine-tuning → Checkpoint engine: trains model
train/train.py - Evaluation datasets → Benchmark converters: uses helpers
evals/datasets/image_policy_mod.py - API caller → Evaluation runner: runs evaluation
README.md - Evaluation runner → Evaluation client: uses client
evals/run.py - Evaluation runner → Metrics: computes metrics
evals/run.py - Evaluation client → Decision service: posts requests
evals/client.py