Loading
Validated training environment: Python3.12.3, Torch2.11+cu130, Transformers5.12.1,
ms-swift4.4.0, Accelerate1.13.0. Loading inference weights does not require the optimizer files.
import torchfrom transformers import AutoProcessor, Qwen3VLForConditionalGeneration repo = "Polly1231/ParaThinker-Qwen3-VL-8B-Instruct-SFT-Qwen3VLPlus-lr4e-5-gbs64-ep1-step153"processor = AutoProcessor.from_pretrained(repo)model = Qwen3VLForConditionalGeneration.from_pretrained( repo, dtype=torch.bfloat16, device_map="auto")
Evaluation recorded so far
Under the original ParaThinker multi-round video evaluation protocol, LVBench has
1,549 completed questions, 385 correct / 24.8547% accuracy, with 538 protocol failures
counted as incorrect. Its native audit passed. These are multi-round protocol results,
not a direct one-shot VQA score. LongVideoBench has no final audited result recorded here yet.
Checkpoint contents and integrity
All 30 original checkpoint files, 135,996,781,789 bytes (126.66 GiB), are preserved:
the eight safetensors shards, processor/tokenizer, complete optimizer and FSDP state,
scheduler, Trainer state and all eight rank RNG states. COMPLETE.json is the original
native checkpoint receipt. NATIVE_FILES_SHA256.json provides the SHA-256 digest and
size of every native file; EXPERIMENT.json records its experiment identity.
No intermediate checkpoint, teacher-data archive, original image or video is included.