Expand description
Top-level Needle SAN (Simple Attention Network) model.
Encoder: N layers of self-attention only (no FFN by default). Decoder: M layers of self-attn + cross-attn (+ optional FFN). Tied embeddings: embedding matrix is shared with LM head (transposed multiply). Inference format: Encoder input: [query_tokens…, tools_id(5), tools_tokens…] Decoder start: [eos_id(1)] Decoder output: <tool_call>(4) then JSON