--- pretty_name: AutoMoT PDM-Lite TransFuser Indexes tags: - automot - autonomous-driving - bench2drive - pdm-lite - transfuser - jsonl --- # AutoMoT PDM-Lite TransFuser Indexes This dataset repo stores the public JSONL indexes used by the AutoMoT PDM-Lite fast-reasoning training path. The files are distributed as `.jsonl.zst` to keep downloads small. Decompress them into `PDM_JSONL_DIR` before training. This repo does not include the full PDM-Lite sensor data or TransFuser++ feature tensors. Each JSONL row points to a precomputed TransFuser++ feature file via `transfuser_feature` and frame id via `transfuser_feature_frame`. ## Files - `pdm_lite_2hz_2tp_train_transfuser.jsonl.zst` - `pdm_lite_2hz_2tp_val_transfuser.jsonl.zst` After decompression, AutoMoT expects: - `pdm_lite_2hz_2tp_train_transfuser.jsonl` - `pdm_lite_2hz_2tp_val_transfuser.jsonl` ## Expected Row Fields Each row keeps the conversation/training metadata and includes: - `transfuser_feature`: relative path to `transfuser_feature/route_features.pt` - `transfuser_feature_frame`: frame id inside the route feature file The prompt uses `` for the TransFuser++ BEV feature token. Front camera images are not fed directly into MoT in this training path. ## Download ```bash huggingface-cli download HqH1111/AutoMoT-PDM-Lite-TransFuser-Indexes pdm_lite_2hz_2tp_train_transfuser.jsonl.zst pdm_lite_2hz_2tp_val_transfuser.jsonl.zst --repo-type dataset --local-dir "$PDM_JSONL_DIR" unzstd -f "$PDM_JSONL_DIR/pdm_lite_2hz_2tp_train_transfuser.jsonl.zst" unzstd -f "$PDM_JSONL_DIR/pdm_lite_2hz_2tp_val_transfuser.jsonl.zst" ``` ## Use With AutoMoT ```bash cd Automot source configs/env.sh CUDA_VISIBLE_DEVICES=0 TOTAL_STEPS=1 DO_EVAL=False NUM_WORKERS=0 bash scripts/train/train_pdm_lite_reasoning.sh ``` The training loader expects the PDM-Lite data root and precomputed TransFuser++ features to be available locally through `PDM_DATA_DIR`.