How erictan7/videoveritas works
VideoVeritas is presented as a video-authenticity detector built around fine-grained perception and fact-based reasoning, with Perception Pretext RL and MintVid as its training and evaluation contributions. The documented user path is deploying the model and submitting a video for inference; a second path covers dataset evaluation and RLHF training. The tree is largely the ms-swift framework, so some VideoVeritas-specific wiring is not identifiable. The sampled engine sources confirm inference backends, but not the model’s detection logic.
Main entry points: infer_engine.py Inference engine · rlhf.py RLHF training · loader.py Dataset loading · vllm_engine.py vLLM backend
Components 17
Video workflow
- Single-video inference
self_scripts/infer/infer_vllm_single.py - Inference commands
swift/cli/infer.py - Model deployment
swift/llm/infer/deploy.py - Detection response
Data and evaluation
- MintVid dataset
- Dataset preparation
swift/llm/dataset/dataset/data_utils.py - Dataset loading
swift/llm/dataset/loader.py - Model evaluation
swift/llm/eval/eval.py
Training
- RLHF commands
swift/cli/rlhf.py - RLHF training
swift/llm/train/rlhf.py - Preference trainers
swift/trainers/rlhf_trainer
Inference runtime
- VideoVeritas model
- Inference engine
swift/llm/infer/infer_engine/infer_engine.py - vLLM backend
swift/llm/infer/infer_engine/vllm_engine.py - PyTorch backend
swift/llm/infer/infer_engine/pt_engine.py - LMDeploy backend
swift/llm/infer/infer_engine/lmdeploy_engine.py
Other components
- Video user
Connections 20
- Video user → Single-video inference: submits video
- Single-video inference → Inference engine: requests inference
- Inference commands → Inference engine: runs inference
- Model deployment → vLLM backend: serves model
- Inference engine → vLLM backend: selects backend
- Inference engine → PyTorch backend: selects backend
- Inference engine → LMDeploy backend: selects backend
- vLLM backend → VideoVeritas model: runs model
- PyTorch backend → VideoVeritas model: runs model
- LMDeploy backend → VideoVeritas model: runs model
- Inference engine → Detection response: returns response
- MintVid dataset → Dataset preparation: provides videos
- Dataset preparation → Dataset loading: prepares data
- Dataset loading → Model evaluation: loads samples
- Video user → Model evaluation: runs benchmark
- Video user → RLHF commands: starts training
- RLHF commands → RLHF training: dispatches training
- RLHF training → Preference trainers: uses trainer
- Dataset loading → RLHF training: supplies training data
- RLHF training → VideoVeritas model: updates model