Modelpublic

yinpei-tri-pi/mira_simplememvla_60k

sourceHugging Faceapache-2.0updated 3d agoView on Hugging Face

0likes

Model Card

SimpleMemVLA RoboMME — 60k-step training run

Evaluation checkpoints from a planned 60,000-step RoboMME training run. The repository currently contains 20,000-step, 30,000-step and 40,000-step snapshots; 50,000-step and 60,000-step snapshots will be added later.

Directory	Optimizer step	Contents
`checkpoint-20000/`	20,000	Model weights and evaluation configuration
`checkpoint-30000/`	30,000	Model weights and evaluation configuration
`checkpoint-40000/`	40,000	Model weights and evaluation configuration

Each checkpoint is self-contained for use with the SimpleMemVLA evaluation code: model.safetensors, config.json, processor_config.json, tokenizer.json, tokenizer_config.json, chat_template.jinja, and stats.json. evaluation_manifest.json records the size and SHA-256 of every payload file. All payloads were verified against the completed training checkpoint before upload.

Optimizer, scheduler, Trainer and RNG state are intentionally omitted. These exports are for evaluation, not exact training-state resume.

Download on another machine

python

from pathlib import Path
from huggingface_hub import snapshot_download

root = snapshot_download(
    repo_id="yinpei-tri-pi/mira_simplememvla_60k",
    allow_patterns=["checkpoint-20000/*", "checkpoint-30000/*", "checkpoint-40000/*"],
    local_dir="mira_simplememvla_60k",
)
checkpoint = Path(root) / "checkpoint-40000"
# Pass this checkpoint directory to the SimpleMemVLA RoboMME evaluator.

Keep each snapshot's stats.json alongside its weights. Use those matching statistics for input normalization and action unnormalization; do not substitute statistics from the official checkpoint or another training run. The tokenizer and processor files must also remain alongside the checkpoint configuration.

Training uses 32 GPUs with per-device batch 16 and gradient accumulation 1 (global batch 512). Video demonstrations are history context, not action targets.