xai-org/x-algorithm · error · ValueError

{mask_name}: seq len {is_candidate.shape[1]} != multi-hot {s

Error message

{mask_name}: seq len {is_candidate.shape[1]} != multi-hot {seq_len}

What it means

The candidate mask's per-row sequence length must equal the sequence length of the multi-hot column (list_size). If the reshaped mask has a different second dimension, the element-wise np.where fold would broadcast incorrectly, so ValueError is raised naming the mask column and both lengths.

Source

Thrown at phoenix/xrex/data/conversion_labels.py:116

        .to_numpy(zero_copy_only=False)
        .astype(np.bool_)
        .reshape(batch.num_rows, seq_len, vocab)
        .copy()
    )
    mask_name = next(
        (n for n in ("newEventMask", "newEventMaskSeq") if n in batch.schema.names), None
    )
    if mask_name is None:
        raise ValueError("batch missing newEventMask; conversion-label folding is candidate-only")
    mask_col = batch.column(mask_name)
    is_candidate = (
        mask_col.flatten()
        .to_numpy(zero_copy_only=False)
        .astype(np.bool_)
        .reshape(batch.num_rows, mask_col.type.list_size)
    )
    if is_candidate.shape[1] != seq_len:
        raise ValueError(f"{mask_name}: seq len {is_candidate.shape[1]} != multi-hot {seq_len}")
    for name in action_cols:
        idx = action_index_of(name)
        if idx >= vocab:
            raise ValueError(f"{name}: bit {idx} outside vocab {vocab}")
        delays_col = batch.column(name)
        delays = (
            delays_col.flatten()
            .to_numpy(zero_copy_only=False)
            .astype(np.int64)
            .reshape(batch.num_rows, delays_col.type.list_size)
        )
        if delays.shape[1] != seq_len:
            raise ValueError(f"{name}: seq len {delays.shape[1]} != multi-hot {seq_len}")
        bits[:, :, idx] = np.where(
            is_candidate, delays_to_labels(delays, window_ms), bits[:, :, idx]
        )
    inner = pa.FixedSizeListArray.from_arrays(pa.array(bits.reshape(-1)), vocab)
    outer = pa.FixedSizeListArray.from_arrays(inner, seq_len)

View on GitHub (pinned to 24c60942c5)

Solutions

1. Regenerate the dataset so mask and multi-hot share the same sequence length.
2. Rebuild the batch with consistent FixedSizeList types before folding.
3. Skip folding for batches whose schema lengths disagree.

Example fix

// before: mask seq 32, multihot seq 64
batch = fold_action_delays_into_multihot(batch, w)  # ValueError

// after: regenerate data with seq_len consistent (e.g. both 64)
batch = fold_action_delays_into_multihot(batch_fixed, w)

seq = batch.column(batch.schema.get_field_index(ACTION_MULTIHOT_COLUMN)).type.list_size
mask = next(n for n in ('newEventMask','newEventMaskSeq') if n in batch.schema.names)
assert batch.column(mask).type.list_size == seq

def seq_lens_consistent(batch) -> bool:
    lens = {batch.schema.field(n).type.list_size for n in batch.schema.names
            if pa.types.is_fixed_size_list(batch.schema.field(n).type)}
    return len(lens) <= 1

try:
    batch = fold_action_delays_into_multihot(batch, w)
except ValueError as e:
    raise RuntimeError(f'bad batch schema: {e}') from e

When it happens

Trigger: Batch where newEventMask has seq len 32 but the multi-hot list_size is 64; schema evolution changing sequence length for one column but not the other; mixing columns from different feature versions into one batch.

Common situations: Mid-migration datasets where sequence length changed; concatenating batches with inconsistent FixedSizeList widths.

Related errors

• {name}: seq len {delays.shape[1]} != multi-hot {seq_len}
• {key}: got {arr.shape}/{arr.dtype}, manifest says {meta['sha
• Tensor {name!r}: checkpoint has shape {tuple(t.shape)}, but
• block-sparse arrays cover {bs_num_blocks} m-tiles but the ke
• {tensor_name}{context_clause} with shape {tensor.shape} cann

AI-assisted analysis of xai-org/x-algorithm@24c60942c5 (2026-08-28). Data as JSON: /api/errors/ae461716acc79593. Report an issue: GitHub.