yujl24/EvoReflex-LIBERO-Decision

EvoReflex LIBERO Decision — ready-to-train data

Canonical Intern-Decision training rows, both RGB cameras, measured robot state, fixed episode splits, and complete training code. The prepared data can train Qwen3.5-0.8B or Qwen3.5-4B directly. The default 4B recipe follows our v3 run.

Contents and coverage

Split	Episodes	Observation/action rows
train	1,340	216,515
calibration	150	23,937
validation	203	33,013
Total	1,693	273,465

The data covers all 40 standard LIBERO tasks: Spatial, Object, Goal and Long, 10 tasks each. Two lossless 256×256 PNGs per row give 546,930 images. Episodes, including all their frames, belong to exactly one split. Calibration is separate from both optimization and checkpoint selection.

Source: lerobot/libero, revision a1aaacb7f6cd6ee5fb43120f673cebb0cfea7dd4, Apache-2.0. These PNGs preserve the decoded RGB of the published AV1 videos. No extra image rotation or crop is applied. SOURCE_DATASET_CARD.md and NOTICE preserve source attribution.

This package contains standard LIBERO demonstrations. LIBERO-Pro is used for separate downstream evaluation, and its evaluation trajectories are not included in training data.

Download and restore

For this private repository, first log in with credentials having read access. For a colleague downloading from this personal repository, the owner can create a fine-grained token with read permission limited to this dataset, and provide it through a secure channel. The colleague enters it at the login prompt below. Credentials are not included in the package.

python3.10 -m venv .venv-download
.venv-download/bin/python -m pip install huggingface-hub==1.17.0
.venv-download/bin/python -c 'from huggingface_hub import login; login()'
.venv-download/bin/python - <<'PY'
from huggingface_hub import hf_hub_download
import shutil
path = hf_hub_download('yujl24/EvoReflex-LIBERO-Decision', 'bootstrap.py', repo_type='dataset')
shutil.copyfile(path, 'bootstrap.py')
PY
.venv-download/bin/python bootstrap.py \
  --repo yujl24/EvoReflex-LIBERO-Decision --destination "$PWD/EvoReflex-LIBERO-Decision"

Use the published commit ID with --revision COMMIT when freezing a run. Downloads resume through the Hub client. SHA-256 checks run before extraction; completed archive receipts allow extraction to continue after interruption. The script preserves the original PNG bytes and prepared metadata hashes.

Keep the downloaded bundle until extraction succeeds. Allow approximately twice the bundle size for download plus extraction, in addition to model/checkpoint storage. The exact byte count is in dataset-manifest.json.

Restored layout:

EvoReflex-LIBERO-Decision/
  code/
    training/jev_libero/       # trainer, compiler, configs, launchers, licenses
    docs/                     # detailed v3/4B and sampling notes
    training-source.json      # exact source commit for a standalone launch
  data/
    libero-decision-nohistory-v2/   # default v3 rows/weights + actual images
    libero-decision-balanced-v4/    # v6 weights/task metadata + relative links
  dataset-ready.json

Both recipes use exactly the same rows, labels, images and splits. The restored v3 directory is independent of the original server. The balanced overlay uses only relative links to its sibling v3 directory; keep both under data/.

Direct 4B training on 8 × 80GB or larger GPUs

cd EvoReflex-LIBERO-Decision/code
python3.10 -m venv .venvs/train
.venvs/train/bin/python -m pip install -r training/jev_libero/requirements.txt
.venvs/train/bin/python -m pip install --no-deps -r training/jev_libero/requirements-fast.txt
.venvs/train/bin/python training/jev_libero/fetch_model.py \
  --config training/jev_libero/config-4b-v3-8gpu.json \
  --destination assets/models/Qwen3.5-4B
export MODEL_PATH="$PWD/assets/models/Qwen3.5-4B"
export DATA_PATH="$(cd ../data/libero-decision-nohistory-v2 && pwd)"
export RUN_PATH="$PWD/runs/jev-libero-qwen35-4b-v3-nohistory-8gpu"
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
mkdir -p logs
set -o pipefail
bash training/jev_libero/launch_4b_v3_8gpu.sh 2>&1 | tee logs/train-4b-v3.log

The launcher verifies prepared files and the fixed base-model download receipt, then starts training. Training does not require a LIBERO simulator. A supported NVIDIA driver, at least 256GiB host memory and roughly 600GiB run storage are recommended for the eight-GPU 4B configuration. Full checkpoint storage is additional to the data package.

The recipe freezes all vision/merger parameters and trains the language backbone/shared LM head. It is full language-parameter training, with optional ZeRO-1 optimizer-state sharding; no LoRA is used. Global batch64, three epochs, LR2e-5, weight decay0.01, and hard-label full-vocabulary CE match v3.

TensorBoard:

.venvs/train/bin/python -m tensorboard.main \
  --logdir "$RUN_PATH/tensorboard" --host 127.0.0.1 --port 18880

Watch train/decision_ce, train/mean_field_accuracy, validation/decision_ce, and validation/mean_field_accuracy; also inspect gripper-switch and task metrics. Select checkpoints using complete validation and compare their closed-loop evaluations. The existing 0.8B v6 validation loss reached its minimum near step2000 and rose later, so the final checkpoint is not automatically the preferred one.

For the eight-GPU 0.8B v3 recipe use launch_no_history8.sh and explicitly set MODEL_PATH, DATA_PATH, and RUN_PATH. For v6 balanced sampling use launch_balanced_v3_8gpu.sh with DATA_PATH pointing to the balanced overlay. The default 4B launcher uses the v3 sampler; it does not silently adopt v6.

Canonical schema and timing

schema/example.json is a complete actual row. The JSONL files contain:

• id: stable episode/frame identity; tracing only.
• images: agentview then wrist, relative to the selected prepared data root.
• state: task, measured robot state, camera descriptions; no action history.
• questions: seven choice questions with frozen criteria order.
• targets: the label plus original continuous action component per field.

Measured 8D state is TCP position3 in meters, orientation3 in axis-angle radians, and left/right finger qpos2 in meters. It is an actual model input. Timing is observation[t] → action[t].

dx/dy/dz/rx/ry/rz each use 17 nonuniform normalized OSC bins from−1 to+1. gripper uses−1 (open/keep open) and+1 (close/keep closed). All seven fields form one simultaneous native action. These bins are controller increments, not centimeter movements. There is no DONE field.

Canonical axis labels are b00–b16, and gripper labels are g0/g1. The pinned Intern-Decision compiler maps them to A–Q and A–B respectively. Supervision uses the LM-head logits at the position immediately before each <decision> marker, with the full vocabulary denominator. The compiler revision is 3572c8a68b5df5dafe02d0e093989ba8ec0183bc; its code/license are included.

At inference, use the checkpoint tokenizer/processor and the same compiler; decode seven legal choices into the native controller vector. Use serve_decision.py/eval_campaign.py for the matching evaluation protocol.

Sampling variants

v3: a frozen 25%/25%/25%/25% mixture of all frames, initial12-frame window, gripper-switch±3-frame window, and motion-change frames. The sampler uses replacement; all rows retain positive probability. Validation is complete and unweighted.

v6 balanced: equal task mass, then equal episode mass within each task, then the same four components. Real switch centers receive score×2; expected draws per row are bounded at4 per epoch. Task IDs and phase flags are metric/sampler metadata, never model inputs. The actual sampler audit accompanies its files.

Package integrity

dataset-manifest.json binds every payload file by size and SHA-256, records the source commit, and lists the exact image-shard count. image-index.jsonl.gz records the relative path, bytes, SHA-256 and shard for every PNG. Original prepared manifests, offsets, phase arrays, weights and episode splits remain unchanged; each original manifest is checked again after restoration. source-provenance.jsonl.gz keeps original episode/frame IDs, timestamps, task IDs and continuous source actions for conversion audits.

Model weights, tokens, user credentials, evaluation videos, and final-test trajectories are not part of this training-data package.