README.md

World Embedding Benchmark

Bidirectional text-video retrieval for PhysicsBench, with faithful LCO-Embedding inference through Transformers and vLLM.

The benchmark uses parsed_text as the default text prompt and the embedded Parquet video field as video input. It reports text-to-video and video-to-text Recall, MRR, and nDCG globally and per family, and can save embeddings, similarity matrices, checkpoints, and family-confusion tables.

Quick start (standard x86_64 CUDA server)

sudo apt-get update
sudo apt-get install -y ffmpeg python3.12-venv
python3.12 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip setuptools wheel
python -m pip install -r requirements.txt
huggingface-cli login

Use official prebuilt PyTorch/vLLM wheels on x86_64. Do not repeat the DGX Spark ARM64 source-build procedure described in ENVIRONMENT.md.

Place datasets under:

datasets/physics-bench-solid-eval
datasets/physics-bench-optics-eval
datasets/physics-bench-fluid-eval
datasets/physics-bench-dynamics-eval

Each family is a directory containing Parquet shards with query_id, case_id, raw_text, parsed_text, and video columns.

Smoke test

python run_retrieval.py \
  --dataset-dir datasets/physics-bench-solid-eval \
  --model lco-omni-3b --model-name ./models/LCO-Embedding-Omni-3B \
  --backend vllm \
  --vllm-max-model-len 32768 --vllm-gpu-memory-utilization 0.7 \
  --video-sampling processor --fps 2 --max-frames 128 \
  --batch-size 4 --video-prefetch-batches 2 \
  --video-decode-workers 8 --video-decoder torchcodec --no-vllm-enforce-eager \
  --text-column parsed_text --limit-videos-per-family 1 \
  --output results/solid_lco_3b_smoke.json

Remove the limit for a full run. On the dual RTX 6000D server, start with batch 4, prefetch 2, and eight TorchCodec decode workers. Run one independent job per GPU; then benchmark larger values on the target server. Registered model keys are lco-omni-3b and lco-omni-7b; --model-name overrides their Hugging Face checkpoints.

Video regression

run_regression.py evaluates a frozen video representation with a deterministic nested cross-validated ridge probe. Outer folds produce out-of-sample predictions; ridge strength is selected only within each outer training fold. It reports MAE, RMSE, R², Pearson and Spearman correlations, and target-range normalized MAE/RMSE.

python run_regression.py \
  --dataset-dir datasets/physics-bench-regression-500 \
  --subset pendulum \
  --model lco-omni-3b --model-name ./models/LCO-Embedding-Omni-3B \
  --backend vllm \
  --vllm-max-model-len 32768 --vllm-gpu-memory-utilization 0.7 \
  --video-sampling processor --fps 2 --max-frames 128 \
  --batch-size 4 --video-prefetch-batches 2 \
  --video-decode-workers 8 --video-decoder torchcodec --no-vllm-enforce-eager \
  --embedding-output-dir results/regression/pendulum/embeddings \
  --output results/regression/pendulum/result.json

The embedding directory makes encoding resumable and stores video_embeddings.npz. Per-example out-of-fold predictions are saved beside the result as predictions.csv.

Fixed test set and scaling protocol

Create the permanent 100-example test split and five nested training orders for all subsets once:

python prepare_regression_splits.py

Manifests under regression_splits/physics-bench-regression-500/ use stable example IDs, ten equal-count target-rank strata, seed 42, training sizes 25, 50, 100, 200, 300, 400, and repetition seeds 1000 through 1004. Existing manifests are validated and are never silently overwritten.

Run a scaling experiment with one video-encoding pass:

python run_regression_scaling.py \
  --subset pendulum \
  --model lco-omni-3b --model-name ./models/LCO-Embedding-Omni-3B \
  --backend vllm --vllm-max-model-len 32768 \
  --vllm-gpu-memory-utilization 0.7 \
  --video-sampling processor --fps 2 --max-frames 128 \
  --batch-size 4 --video-prefetch-batches 2 \
  --video-decode-workers 8 --video-decoder torchcodec --no-vllm-enforce-eager \
  --embedding-output-dir results/regression-scaling/pendulum/embeddings \
  --output results/regression-scaling/pendulum/result.json

Every size is tuned only on its selected training subset. The same fixed 100 examples are then used for test metrics across all sizes and repetitions.

Fidelity and sampling

• Processor mode reproduces Qwen Omni FPS sampling while decoding only selected frames; fixed mode uniformly selects exactly --num-frames N frames.
• Both backends preserve LCO's compression prompts, LAST-token pooling, and L2 normalization.
• compare_lco_backends.py validates Transformers/vLLM embedding parity.

Main files

• run_retrieval.py: evaluation CLI.
• run_regression.py: video-representation regression CLI.
• run_regression_scaling.py: fixed-test regression scaling CLI.
• prepare_regression_splits.py: deterministic split-manifest generator.
• run_retrieval.sh: readable full-run examples.
• world_embedding_benchmark/retrieval.py: loading, caching, scoring, metrics.
• world_embedding_benchmark/regression.py: regression loading, probing, and metrics.
• world_embedding_benchmark/models/: Transformers/vLLM LCO adapters.
• world_embedding_benchmark/embedding_artifacts.py: resumable artifacts.
• compare_lco_backends.py: backend parity.
• benchmark_lco_video_throughput.py: throughput tuning.
• debug_lco_retrieval.py: retrieval diagnostics.
• visualize_similarity_matrices.py: embedding visualization.
• merge_fluid_eval_captions.py: legacy fluid reconstruction.
• HANDOFF.md: project state and next steps.
• ENVIRONMENT.md: x86_64 setup and historical ARM64 notes.

Datasets, weights, results, caches, environments, and local build trees are intentionally excluded from Git.