RAVEC-7B/README.md

RAVEC-7B: main-result checkpoint

This is the M1 step100 Actor export used for RAVEC Table 2 (Avg 61.00), derived from Qwen/Qwen2.5-VL-7B-Instruct.

All 15 original model files match the SHA-256 hashes recorded in all 14 evaluation runs. No conversion or quantization was applied. See checkpoint_manifest.json for the exact source, file hashes and ten benchmark scores; evaluation_summary.json retains the original evaluation summary. The reported ten-benchmark average differs from the descriptive average across all 14 evaluation tasks.

from transformers import AutoProcessor, Qwen2_5_VLForConditionalGeneration
repo = "Jack04810/RAVEC"
processor = AutoProcessor.from_pretrained(repo, subfolder="RAVEC-7B")
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    repo, subfolder="RAVEC-7B", torch_dtype="auto", device_map="auto"
)

The original export records Transformers 4.51.3. Original configuration and tensor files are preserved unchanged.

License

The 7B checkpoint derives from Qwen2.5-VL-7B-Instruct under Apache-2.0. The base model card declares Apache-2.0; the accompanying LICENSE contains the official Apache-2.0 text. This is a fine-tuned derivative released as RAVEC; its weights differ from the base model.