chocopan-oft-track3-ckpt-v2

chocopan-oft-track3-ckpt-v2

Periodic training checkpoints from a LoRA fine-tune of Sylvest/openvla-7b-oft-finetuned-libero-plus-mixdata on a mixture of synthetic reverse-manipulation demonstrations and forward LIBERO-plus demonstrations. This repository is a checkpoint archive, not a ready-to-load model: each step_<N>/ directory holds the LoRA adapter and the two trained heads, and must be merged into the base model before it can be used.

This is the -v2 run: a shorter schedule with a lower learning rate, a short warmup and a step decay.

track3 in the repository name is just an internal tag for the data this run was trained on; see Training data for what it actually is.

What is trained

The base model is an OpenVLA-7B OFT policy already fine-tuned on LIBERO-plus. This run continues from its absolute step 150,000 and trains:

• a LoRA adapter (rank 32, dropout 0.0, all-linear target modules) over the VLA backbone;
• the L1 regression action head (all parameters);
• the proprio projector (all parameters).

The interface is unchanged from the base recipe:

libero_10 is the right key for every checkpoint here: the fine-tuning datasets were seeded with the base checkpoint's statistics, so every entry in dataset_statistics.json carries the base model's numbers.

Repository layout

checkpoints/
  step_150100/
    lora_adapter/
      adapter_config.json
      adapter_model.safetensors
      README.md
    action_head--150100_checkpoint.pt
    proprio_projector--150100_checkpoint.pt
    dataset_statistics.json
    preprocessor_config.json
    processor_config.json
    processing_prismatic.py
    tokenizer.json
    tokenizer.model
    tokenizer_config.json
    added_tokens.json
    special_tokens_map.json
  step_150200/
  ...
  step_153000/

Checkpoints are written every 100 absolute steps. As of 2026-09-10 the archive holds 30 of them, step_150100 through step_153000; browse the file list for the current set. No merged backbone weights are stored in any directory -- there is no config.json and no model*.safetensors until you produce them in step 3 below.

Usage

1. Pin the code

Inference needs the OpenVLA-OFT fork of transformers -- the checkpoint's modeling_prismatic.py is loaded with trust_remote_code=True and is written against it -- and merging needs the OpenVLA-OFT scripts. Both are pinned:

git clone https://github.com/moojink/openvla-oft.git
git -C openvla-oft checkout e4287e94541f459edc4feabc4e181f537cd569a8
git clone https://github.com/moojink/transformers-openvla-oft.git
git -C transformers-openvla-oft checkout bc339d9ad707454c0c115970db43c260067c61ab
pip install --no-deps -e ./openvla-oft -e ./transformers-openvla-oft

--no-deps is deliberate: the OpenVLA-OFT pyproject.toml pins torch==2.2.0, which you will usually want to ignore in favour of a build that matches your GPU.

2. Download the base model and one checkpoint

hf download Sylvest/openvla-7b-oft-finetuned-libero-plus-mixdata \
  --revision a85655ec941bae6644c9fbdf62db02b9726d7cf5 --local-dir ./base
hf download chocopan/chocopan-oft-track3-ckpt-v2 --include "checkpoints/step_153000/*" --local-dir ./ckpt

3. Merge the LoRA adapter into the base weights

checkpoints/step_<N>/ holds only the adapter and the two heads, so the merged backbone has to be produced once. The script writes config.json and the merged *.safetensors into the checkpoint directory, and it needs a CUDA GPU (it moves the model to cuda).

python openvla-oft/vla-scripts/merge_lora_weights_and_save.py \
  --base_checkpoint ./base \
  --lora_finetuned_checkpoint_dir ./ckpt/checkpoints/step_153000

4. Run the policy

import json
from pathlib import Path

import torch
from transformers import AutoModelForVision2Seq, AutoProcessor
from prismatic.models.action_heads import L1RegressionActionHead
from prismatic.models.projectors import ProprioProjector

MODEL_DIR = Path("./ckpt/checkpoints/step_153000")   # after the merge in step 3
STEP = 153000
DTYPE = torch.bfloat16


def load_state(path):
    state = torch.load(path, map_location="cpu", weights_only=True)
    return {(k[7:] if k.startswith("module.") else k): v for k, v in state.items()}


model = AutoModelForVision2Seq.from_pretrained(
    MODEL_DIR, trust_remote_code=True, torch_dtype=DTYPE, low_cpu_mem_usage=True
)
model.vision_backbone.set_num_images_in_input(2)     # third-person + wrist
model = model.to("cuda").eval()
processor = AutoProcessor.from_pretrained(MODEL_DIR, trust_remote_code=True)

action_head = L1RegressionActionHead(
    input_dim=model.llm_dim, hidden_dim=model.llm_dim, action_dim=7
)
action_head.load_state_dict(load_state(MODEL_DIR / f"action_head--{STEP}_checkpoint.pt"))
action_head = action_head.to("cuda", dtype=DTYPE).eval()

proprio_projector = ProprioProjector(llm_dim=model.llm_dim, proprio_dim=8)
proprio_projector.load_state_dict(
    load_state(MODEL_DIR / f"proprio_projector--{STEP}_checkpoint.pt")
)
proprio_projector = proprio_projector.to("cuda", dtype=DTYPE).eval()

model.norm_stats = json.loads((MODEL_DIR / "dataset_statistics.json").read_text())
UNNORM_KEY = "libero_10"

# third_person / wrist: 224x224 RGB PIL images in the OpenVLA-OFT LIBERO convention.
# proprio: float32 (8,) = [eef_xyz(3), eef_axis_angle(3), gripper_qpos(2)], normalised to
#          [-1, 1] with the q01/q99 in model.norm_stats[UNNORM_KEY]["proprio"].
prompt = f"In: What action should the robot take to {instruction.strip().lower()}?\nOut:"
inputs = processor(prompt, third_person)
wrist_inputs = processor(prompt, wrist)
inputs["pixel_values"] = torch.cat([inputs["pixel_values"], wrist_inputs["pixel_values"]], dim=1)
inputs = {
    k: (v.to("cuda", DTYPE) if torch.is_tensor(v) and v.is_floating_point()
        else v.to("cuda") if torch.is_tensor(v) else v)
    for k, v in inputs.items()
}

with torch.inference_mode():
    actions, _ = model.predict_action(
        **inputs,
        unnorm_key=UNNORM_KEY,
        do_sample=False,
        proprio=proprio,
        proprio_projector=proprio_projector,
        noisy_action_projector=None,
        action_head=action_head,
        use_film=False,
    )
# actions: (8, 7) -- eight consecutive end-effector actions

Image and action conventions follow OpenVLA-OFT's LIBERO helpers: both camera images are rotated 180 degrees before being resized to 224x224, and the predicted gripper channel is mapped from [0, 1] (0 = close) to the simulator's +1 = close convention.

Training configuration

In upstream finetune.py, --lr_warmup_steps > 0 rewrites the learning rate on every gradient step and therefore cancels the --num_steps_before_decay milestone. This run patched the warmup block to apply only while gradient_step_idx < lr_warmup_steps, so warmup and decay coexist.

Reproduction (OpenVLA-OFT vla-scripts/finetune.py at e4287e94541f459edc4feabc4e181f537cd569a8; --resume_step, --max_steps and --save_freq are absolute step numbers):

torchrun --standalone --nnodes 1 --nproc-per-node 1 vla-scripts/finetune.py \
  --vla_path <base checkpoint> --resume True --resume_step 150000 \
  --data_root_dir <dir holding the RLDS datasets> --dataset_name <registered mixture> \
  --run_root_dir <output dir> \
  --use_l1_regression True --num_images_in_input 2 --use_proprio True --image_aug True \
  --use_lora True --lora_rank 32 --lora_dropout 0.0 \
  --batch_size 8 --grad_accumulation_steps 4 \
  --learning_rate 5e-5 --lr_warmup_steps 200 --num_steps_before_decay 2000 \
  --max_steps 153000 --save_freq 100 --merge_lora_during_training False

Note that --resume True loads merged weights plus the two heads from --vla_path and re-initialises the LoRA adapter; lora_adapter/ is not read back. Resuming across sessions therefore requires a merged directory, which is what vla-scripts/merge_lora_weights_and_save.py produces.

Training data

Fine-tuning used a mixture of three RLDS datasets, each registered with sampling weight 1.0. OpenVLA-OFT multiplies the weight by each dataset's transition count, so the effective sampling ratio is the transition ratio:

The two reverse datasets are scripted-oracle demonstrations of the inverse of LIBERO / LIBERO-plus pick-and-place tasks: putting an object that sits in a container or on a plate back on the table. The forward dataset is a subset of lerobot/libero_plus (revision f3f49f426d75030177b18778374005bc12ccd588).

Limitations and notes

• This repository stores intermediate training checkpoints, not a selected final model. Nothing here says which step behaves best; choose a step by evaluating it yourself.
• Merged backbone weights are not stored, so a checkpoint is unusable until it has been merged into the base model (see Usage).
• The action / proprio normalisation statistics are the base checkpoint's, injected into the fine-tuning datasets before training. Read them from the checkpoint's dataset_statistics.json and use unnorm_key="libero_10"; recomputing statistics from the fine-tuning data would change the action scale.
• The policy was trained and exercised only in simulation (LIBERO / LIBERO-plus scenes, OSC_POSE control at 20 Hz). It has never been run on physical hardware.
• Most of the fine-tuning signal comes from synthetic, scripted-oracle demonstrations rather than human teleoperation. The motion style is the generator's, and closed-loop behaviour away from the 45 generated task variants can be unreliable.

Sources and license

The contents of this repository are released under the MIT license. Upstream terms still apply to the base weights and to anything derived from LIBERO / LIBERO-plus; the LIBERO-plus source repository carries no license file, while its Hugging Face distribution is published as MIT.