FBN.2025.04
AskVid-Pro
Machine learning engineer, sole build
Video-LLaMA2 (7B), QLoRA (4-bit), OpenCLIP, FAISS, 2025
<10s
to a timestamped answer
7B
parameters, fine-tuned
4-bit
QLoRA on a laptop RTX 3050
2
benchmarks: MSRVTT-QA and NExT-QA
The problem
Answering a question about a video usually means watching the video. Existing question-answering systems are slow, uncited and difficult to trust.
The approach
A 7-billion-parameter Video-LLaMA2, fine-tuned with 4-bit QLoRA on a single laptop RTX 3050, sits behind an OpenCLIP retrieval layer indexed in FAISS, so the model reasons over the moments that matter rather than every frame. ViQAgent-style multi-step reasoning lifts zero-shot accuracy, the pipeline is evaluated on MSRVTT-QA and NExT-QA, and a Streamlit interface on Hugging Face Spaces serves it to anyone.
The outcome
Free-form questions about short videos, answered in under ten seconds, each answer carrying timestamps a sceptic can check.
What it does
4 capabilitiesAnswers with receipts
Every answer points to the exact seconds of footage that support it. Verifiable, never vibes.
A seven-billion-parameter model on a laptop
4-bit QLoRA brought a 7B fine-tune within reach of consumer VRAM, without surrendering the model's size.
Retrieval before reasoning
OpenCLIP embeddings in a FAISS index hand the model only the moments that matter.
Open to anyone
Upload a clip, ask a question, receive a cited answer, all through a Streamlit app on Hugging Face Spaces.
The strata
How it standsIngestion
Frame sampling from short videos under five minutes
Retrieval
OpenCLIP embeddings indexed in FAISS for moment lookup
Reasoning
Video-LLaMA2 7B (4-bit QLoRA) with ViQAgent-style multi-step reasoning
Evaluation
MSRVTT-QA and NExT-QA benchmark suite
Serving
Streamlit application on Hugging Face Spaces
The judgments
The reasoning, on the recordChose
4-bit QLoRA
over full fine-tuning
The only GPU available was a laptop RTX 3050. Quantised adapters made a 7B fine-tune fit in consumer VRAM without giving up the model's size.
Chose
Retrieval before reasoning
over feeding the model every frame
OpenCLIP and FAISS hand the model only the moments that matter, which keeps answers under ten seconds and gives each one a timestamp to cite.
Next exhibit — V
ScopeCompose