home/writing/prompt-seq2seq-design

Prompt Seq2seq Design

GPTClaudeGemini··▲ 372 copies·updated 2026-07-14

prompt-seq2seq-design.prompt

---
name: seq2seq-design
description: Design a sequence-to-sequence pipeline for a given task.
phase: 5
lesson: 09
---

Given a task (translation, summarization, paraphrase, question rewrite), output:

1. Architecture. Pretrained transformer encoder-decoder (BART, T5, mBART, NLLB) is the default. RNN-based seq2seq only for specific constraints (streaming, edge inference, pedagogy).
2. Starting checkpoint. Name it (`facebook/bart-base`, `google/flan-t5-base`, `facebook/nllb-200-distilled-600M`). Match checkpoint to task and language coverage.
3. Decoding strategy. Greedy for deterministic output, beam search (width 4-5) for quality, sampling with temperature for diversity. One sentence justification.
4. One failure mode to verify before shipping. Exposure bias manifests as generation drift on longer outputs; sample 20 outputs at 90th-percentile length and eyeball.

Refuse to recommend training a seq2seq from scratch for under ~1M parallel examples. Flag any pipeline using greedy decoding for user-facing content as fragile (greedy repeats and loops).

when to use it

Community prompt sourced from the open-source GitHub repo DipakMandlik/AIByDM (MIT). A "Prompt Seq2seq Design" style prompt — adapt the placeholders and specifics to your task. Imported as-is and not independently retested here, so check the output before relying on it.