CausalLens (Qwen3-VL-8B)

Weights for CausalLens: a Qwen3-VL-8B model with a LoRA adapter on the language model and a causal residual Q-Former on the visual tokens. The Q-Former applies a backdoor-adjustment style intervention over a fixed confounder dictionary and adds a small, gated residual to every visual token before it enters the language model.

Code and usage: https://github.com/litcoderr/CausalLens

Files

Path	Content
adapter_config.json, adapter_model.safetensors	LoRA adapter (r=16, alpha=32; q/k/v/o/gate/up/down projections of the LLM)
causal_qformer_adapter.pt, causal_qformer_config.json	Causal residual Q-Former (weights + confounder dictionary 16384x256 + prior)
tokenizer / preprocessor_config.json / chat_template.jinja	Processor used for evaluation
base/	Qwen3-VL-8B base checkpoint that the adapters were trained on (fine-tuned from Qwen/Qwen3-VL-8B-Instruct, weights merged)

Usage

git clone https://github.com/litcoderr/CausalLens.git && cd CausalLens
./setup_env.sh                  # Python 3.11 venv with pinned dependencies
./download.sh                   # this repository -> checkpoints/CausalLens (base -> checkpoints/CausalLens/base)
./run_mmvet.sh                  # MM-Vet inference (data setup: see the GitHub README)

From Python (inside the CausalLens repository):

from causallens import load_model_and_processor

model, processor = load_model_and_processor(
    base_model="checkpoints/CausalLens/base",
    checkpoint="checkpoints/CausalLens",
)