Veyra-50M model.safetensors appears to contain the model card instead of model weights
Hi,
I’m trying to test Veyra-50M and I encountered what appears to be an issue with the published model.safetensors file.
Loading the model with:
AutoModelForSequenceClassification.from_pretrained("Dl26/Veyra-50M")
fails with:
SafetensorError: Error while deserializing header: header too large
I inspected the downloaded model.safetensors file directly. Instead of containing valid safetensors model weights, its contents begin with:
license: apache-2.0
language:
pipeline_tag: text-classification
which appears to be the YAML/Markdown content of the model card.
I also downloaded model.safetensors manually through the Hugging Face website to rule out a local cache or Transformers download issue, and obtained the same result.
The file I receive is:
Size: 191,279,348 bytes
SHA-256: b54b43e0ecc3ec40596e2014c154077c4eec1dc89e233ff6e3b014406a2fbad4
The tokenizer and configuration load correctly; the failure occurs when loading the model weights.
Could you please check whether the correct Veyra-50M weights were uploaded to the repository and, if necessary, re-upload the valid model.safetensors checkpoint?
I’m very interested in benchmarking Veyra-50M for an English intent-classification project, so I’d be happy to test the corrected checkpoint.
Thank you!
Hey,
Sorry for responding late, as we aren't very active here.
We searched our sandbox for the missing model files, but we were only able to find the Veyra 100M and 20M weights. Unfortunately, we don't have the weight's for the 50M Veyra model