AlphaBrainGroup/llamaoft-cl-lora-libero-goal

LlamaOFT-CL-LoRA (LIBERO-Goal)

LoRA continual-learning checkpoint on top of an 11 B vision-language backbone, released with the AlphaBrain framework. Provided for direct download and evaluation — no retraining needed.

A LlamaOFT Vision-Language-Action (VLA) model — Llama-3.2-11B-Vision backbone feeding a lightweight MLP action head — fine-tuned sequentially over the 10 LIBERO-Goal tasks with Low-Rank Adaptation (LoRA, r=16) on the VLM and Experience Replay (ER, buffer 1000/task) to mitigate catastrophic forgetting. Ships the final task checkpoint only.

Overview

Architecture	LlamaOFT (Llama-3.2-11B-Vision + MLP action head)
Base VLM	meta-llama/Llama-3.2-11B-Vision-Instruct
Parameters	~11 B total · ~200 M trainable
LoRA	r = 16, α = 16, dropout 0.05, target_modules: all-linear
Continual-learning	Experience Replay, buffer 1000/task, replay ratio 0.5
Task stream	LIBERO-Goal · 10 tasks · 5 000 steps/task

Results

Metric	Value
Average Success Rate (Avg SR)	~17 %
Negative Backward Transfer (NBT, ↑ better)	+0.50
Naive sequential fine-tuning baseline (no ER)	< 10 %

Numbers are conservative estimates over our internal runs; per-run variance is a few percentage points. The larger backbone paired with a lower LoRA rank and a smaller step budget yields a lower Avg SR than the 3 B-scale variants, but the NBT margin shows Experience Replay still provides clear forgetting mitigation over the naive baseline. Reproduction numbers higher or lower than reported are expected — please file an issue / PR with details.

Files

├── README.md                                       model card
├── config.yaml                                     training config (OmegaConf)
├── dataset_statistics.json                         action normalisation (required for inference)
├── task_9_id9_steps_50000_lora_adapter/            LoRA adapter weights + config
└── task_9_id9_steps_50000_action_model.pt          non-VLM weights (MLP action head)

Usage

git clone https://github.com/AlphaBrainGroup/AlphaBrain.git
cd VLA-Engine-Developer
pip install -e .

export PRETRAINED_MODELS_DIR=/path/to/models   # must contain Llama-3.2-11B-Vision-Instruct/

huggingface-cli download AlphaBrainGroup/llamaoft-cl-lora-libero-goal \
    --local-dir ./llamaoft_cl_lora

python -m AlphaBrain.training.trainer_utils.peft.merge_lora_checkpoint \
    --base_config      configs/continual_learning/llamaoft_cl_lora_libero.yaml \
    --lora_adapter_dir ./llamaoft_cl_lora/task_9_id9_steps_50000_lora_adapter \
    --action_model_pt  ./llamaoft_cl_lora/task_9_id9_steps_50000_action_model.pt \
    --output_path      ./llamaoft_cl_lora_final.pt

python deployment/model_server/server_policy.py \
    --ckpt_path ./llamaoft_cl_lora_final.pt --port 10093 --use_bf16

Reproduction

bash scripts/run_continual_learning_scripts/run_cl_train.sh \
    --yaml configs/continual_learning/llamaoft_cl_lora_libero.yaml

License

MIT for the AlphaBrain-specific artifacts (adapter, action head, configs). The base Llama-3.2 weights are distributed under Meta's Llama 3.2 Community License and must be accepted separately before downloading the base model.

Citation

@misc{alphabrain2026,
  title  = {AlphaBrain: A Modular Open-Source Framework for Embodied Intelligence Research},
  author = {AlphaBrain Team},
  year   = {2026},
  url    = {https://github.com/AlphaBrainGroup/AlphaBrain}
}