TrackAnything — midtraining

LargeMix4M-Ref5, V7 neural memory reader + V8 external memory bank. This revision exports Stage2 step 59,375 / 59,375, following 3,125 Stage1 updates: 8,000,000 / 8,000,000 sample exposures. Status: completed two-epoch midtraining. Pin the Hub commit revision to reproduce this snapshot.

The four-million-row manifest mixes VastTrack 50%, LaSOT 25%, TNL2K 15%, TNLLT 5%, and RefCOCO 5%; epoch two reshuffles the same manifest. Training uses six memory slots x eight tokens, rank-16 LLM LoRA, frozen visual backbone/base LLM, NTP + MTP, and auxiliary continuous bbox/presence losses. Stage1 is 0.1 epoch of memory/head/query warmup; Stage2 is 1.9 epochs with LLM LoRA and visual projection enabled. This is the checkpoint before the subsequent TNL2K RobustMix20 posttraining.

Load

Use the TrackAnything runtime and its documented environment (source scripts/env.sh).

from huggingface_hub import snapshot_download
from tracking_anything.model_release import load_released_model, verify_release
root = snapshot_download("trackanything-hkust-loongforge/midtraining")  # Set revision to a Hub commit.
verify_release(root)
model, processor, metadata = load_released_model(root, device="cuda")

This is the project's base + unmerged FP32 overlay bundle, loaded with the project runtime rather than vanilla AutoModel. The exact base, tokenizer, 622 trained tensors / 366,832,647 parameters, licenses, and SHA256 manifest are included. Trained tensors are checked bit-for-bit against the source. Optimizer/RNG states, training manifests and images are excluded. This bundle supports inference and weight initialization, not exact optimizer resume. Bundled processor Python code requires trust in the source.

No full TNL2K tracking benchmark is claimed for this revision. Validation curves from training do not establish online coupled tracking performance. Weights retain the upstream NVIDIA model license; project code is licensed separately. Models are improved using Qwen.