FLEX34-PureLoRA

Final PureLoRA checkpoint after all 34 FLEX tasks, containing every trained expert and the NCM routing prototypes. The default router uses softmax logit scale 28 and equal image/text weights. Different routers can reuse these same experts without retraining.

• Code and environment setup
• FLEX instructions and image sources
• Task-to-expert mapping and per-task Base results

The Base reference scores are M = 54.60, G = 60.81%, and R = 83.77%. This is a custom PURELORA adapter: use the FLEX repository's bundled PEFT implementation and loading code. Download the base model and CLIP separately.

Download

hf download yhyet/FLEX34-PureLoRA --local-dir ./checkpoints/FLEX34-PureLoRA

The checkpoint consists of adapter_model.safetensors, adapter_config.json, config.json, and purelora_state.pt. Keep all four in the same directory. To verify the downloaded files, run sha256sum -c SHA256SUMS in that directory.

Inference

Install the FLEX code and prepare the instructions and images following its README. In config/paths/llava_paths.py, set BASE_MODEL_PATH to your local LLaVA-1.5-7B snapshot and CLIP_PATH to your local CLIP ViT-L/14-336 snapshot.

From the FLEX code directory, after downloading this checkpoint there:

export FLEX_INSTRUCTION_ROOT=/path/to/datasets/FLEX/instructions
export PURELORA_MIXTURE_LOGIT_SCALE=28
export PURELORA_IMAGE_WEIGHT=0.5
export PURELORA_TEXT_WEIGHT=0.5

python run.py infer $(seq 0 33) \
  --benchmark flex --method purelora --backbone llava \
  --model-path ./checkpoints/FLEX34-PureLoRA \
  --checkpoint-task 33 --cl-task-num 34 --stage last \
  --gpus 0 --batch-size 1 --temperature 0 --conv-mode vicuna_v1 \
  --no-top1-expert-routing --no-oracle-expert-routing \
  --accel-eval --run-stamp flex34-purelora-ncm28

For a short run, replace $(seq 0 33) with 0. The command uses the FLEX code's default bfloat16 inference. For task-oracle inference, replace --no-oracle-expert-routing with --oracle-expert-routing.

Base model revision: 4481d270cc22fd5c4d1bb5df129622006ccd9234. Expert IDs 0–33 follow the fixed FLEX task order in task_mapping.json. Use of the adapters is subject to the upstream base model and dataset terms.