0step test
0step-test/ contains the unmodified official initial Qwen3.5-9B assets from
revision c202236235762e1c871ad0ccb60c8ee5ba337b9a.
Global step 0, zero optimizer updates, zero training samples consumed.
This is an upload/connectivity test, not a trained model or full training-resume checkpoint.
The initial assets were copied on the Hub using its official cross-repository copy API;
test metadata was uploaded from the archive client.
Checkpoint policy
Save full server checkpoints at steps 2000, 4000, 6000, 10000, 14000, 18000, ...
After step 6000, the scheduled interval is 4000 attempted global training windows.
A coordinated user stop or clean training end also saves its final full state.
ALL full server checkpoints are retained. Neither Trainer checkpoint rotation nor a
successful model-weight upload deletes any server checkpoint. Each complete checkpoint
independently contains model weights, FP32 master partitions, Adam, scheduler, RNG,
Trainer state, quarantine state and the epoch/data cursor needed for exact continuation.
Select any retained checkpoint using an explicit --resume_from_checkpoint path;
auto selects the newest compatible complete checkpoint. The dataset index, world size,
batch, seed, preprocessing and runtime must match the selected checkpoint.
Upload only model safetensors plus model/generation config, tokenizer and processor
assets to weights/checkpoint-STEP/. Do not upload Adam, FP32 master partitions,
scheduler, RNG, Trainer state, training arguments or sampler/data-cursor files.
Hub weights support inference or a fresh fine-tune; exact optimizer and data-cursor
resume uses the full server checkpoint.
Uploads verify remote sizes and content hashes before writing an acknowledgement.
Only the regenerable Mac upload cache is removed after success; server checkpoint
retention is unchanged. Pending or failed uploads retry independently of training.
Hub history is not automatically purged, and cumulative private weights consume
account storage. Formal continuous training was started on 2026-10-30 (Asia/Shanghai).
Active training configuration
Qwen/Qwen3.5-9B, full-parameter BF16 training on 8 nodes / 64 GPUs.
Per-GPU batch 8, gradient accumulation 1, global batch 512.
ZeRO-2 with overlap_comm=false, gradient checkpointing, max_length=4096,
min_pixels=4096, max_pixels=262144. AdamW LR 2e-5, 1000 warmup windows,
constant LR afterward; continuous across epochs until the user stops it.
The pinned SpatialLogic data revision is 82e5aae49f61799e5aff317e7548ad8395b648bc.
Global input/target deduplication removed 40,253 rows. One source archive was
verified to match its official SHA256 but was itself truncated; 85 rows referencing
unavailable images were excluded. All complete original data remains on the server.
The fixed training index has 27,543,089 rows and 53,795 global windows per epoch;
49 trailing rows are dropped per epoch instead of padding with repeated samples.
Validation videos/episodes are disjoint from training, including shared agibot
identities across short/long/leap sources. A fixed 1,792-row held-out evaluation
sample is used every 200 windows.
Scheduled trained snapshots are queued for upload only after the full server
checkpoint is committed. The existing 0step-test folder remains an untrained
initial-model connectivity test. No newly trained snapshot is claimed uploaded
until its remote size and hashes are verified.