Gene-token deletion and rank promotion

Your task

Function signature

gene_token_intervention(
    tokens: torch.Tensor,
    valid_mask: torch.Tensor,
    protected_mask: torch.Tensor,
    target_ids: torch.Tensor,
    target_mask: torch.Tensor,
    mode: str = 'delete',
    pad_token_id: int = 0,
) -> dict[str, torch.Tensor]

Worked example

Promotion keeps identity and provenance

Parameters

Return

dict[str, torch.Tensor]

Required behavior

• Inputs are tokens [B,L] int64 with B,L\ge 1, prefix-contiguous valid_mask [B,L], protected_mask [B,L] (true only at valid boundary special tokens), target_ids [B,K] int64 and target_mask [B,K] with K\ge 1. Valid token IDs are unique within a cell, nonnegative and not pad_token_id.
• Active target IDs obey the same rule and are unique within a cell. Invalid token and inactive target payloads are ignored. Protected positions may form a prefix and/or suffix of the valid sequence, never a protected token between two mutable genes. If every valid token is protected, treat the entire protected block as a prefix.
• For mode='delete' (default), remove every mutable gene whose ID is an active target. Missing targets have no effect.
• For mode='promote', emit all active, nonprotected target IDs at the front of the MUTABLE gene region, in supplied target order: move an existing gene there once or insert a missing gene. Follow these with the nontarget mutable genes in their original order.
• In either mode, a target equal to a protected token is ignored for editing; that protected token remains untouched. Keep the protected prefix, the first L-prefix_length-suffix_length edited mutable entries, and the protected suffix. Thus truncation can drop mutable tail genes but never a protected suffix.
• Pad the remaining tail to L with pad_token_id and false validity.
• Return tokens [B,L], valid_mask [B,L], lengths [B], source_positions [B,L] (original position for each retained/moved token; -1 for an inserted or padded entry), intervention_mask [B,L] (true only on a retained promoted/inserted target), and target_found [B,K] (true iff that ACTIVE target occurred anywhere in the original valid sequence, including a protected token).
• All integer outputs are int64 and masks Boolean.
• For delete the intervention mask is all false. Empty cells and deleting every gene are supported. No mutation of inputs. Config: mode is 'delete' or 'promote'; pad_token_id defaults to 0 and must be a nonnegative integer.
• Unsupported mode, duplicate/invalid valid IDs or active targets, non-prefix validity, protected padding or interior protected tokens raise ValueError. V1 ordinarily supplies no boundary specials: protected_mask is all false.
• Optional protected boundaries are an explicit safety extension, not V2 tokenization parity, and no special tokens are automatically inserted.

Why this matters

Starter code

You write the body of this function in the BioTorch workspace. Your code runs in an isolated PyTorch sandbox; Run checks the public examples and Submit grades the full test suite, including gradients.

import torch


def gene_token_intervention(
    tokens: torch.Tensor,  # (B, L)
    valid_mask: torch.Tensor,  # (B, L)
    protected_mask: torch.Tensor,  # (B, L)
    target_ids: torch.Tensor,  # (B, K)
    target_mask: torch.Tensor,  # (B, K)
    mode: str = 'delete',
    pad_token_id: int = 0,
) -> dict[str, torch.Tensor]:
    """Return dict[str, torch.Tensor]:
    
    tokens: (B, L).
    source_positions: (B, L).
    intervention_mask: (B, L). Always returned; all False in delete mode.
    valid_mask: (B, L).
    lengths: (B,).
    target_found: (B, K)."""
    # Implement the operation. Inputs and test setup are supplied by BioTorch.
    raise NotImplementedError("Implement gene_token_intervention")

Related exercises

• Builds onDelete Gene Tokens
• Builds onPromote Gene Tokens

Sources and further reading

• www.nature.com
• huggingface.co
• huggingface.co