This line is too permissive; I agree that the previous logic was too strict, but I prefer a middleground:

• if unk_token is absent: unk_token_id = None
• if unk_token is present but not found in vocab: error
• if unk_token is present and in vocab: use the unk_token

So something like this should work I think:

let unk_token_id = match unk_token {
    None => None, // Allow None if tokenizer does not declare one
    Some(tok) => match tokenizer.token_to_id(tok) {
        Some(id) => Some(id as usize),
        None => {
            return Err(anyhow!(
                "tokenizer declares unk_token='{tok}' but it isn't in the vocab"
            ))
        }
    },
};

This does an extra full copy of the embeddings which adds up for larger models. I think the easiest solution is to change embeddings in from_raw_parts from embeddings: &[f32, to embeddings: Vec[f32] and then this becomes:

let embeddings = Array2::from_shape_vec((rows, cols), embeddings)
    .context("failed to build embeddings array")?;

And then Self::from_raw_parts(tokenizer, &floats, rows, cols, normalize, weights, token_mapping) becomes Self::from_raw_parts(tokenizer, floats, rows, cols, normalize, weights, token_mapping). This way from_pretained can copy the embeddings over directly.