Skip to content

pe_rewrite.md

Latest commit

 

History

History
127 lines (95 loc) · 6.77 KB

pe_rewrite.md

File metadata and controls

127 lines (95 loc) · 6.77 KB
Raw

PE Rewrite Distillation

The pe_rewrite_distill pipeline distills a multi-step prompt-rewriting agent (plan -> rewrite -> reflection) into training data for a lightweight student model. Instead of answering questions, the student learns to rewrite short text-to-image prompts into detail-rich, production-ready prompts in a single call.

For the JSONL schemas used by each stage, see data_formats.md. 中文版见 pe_rewrite_zh.md。

When to use this pipeline

Use this pipeline when you want to go from a set of raw text-to-image prompts (or a small seed set) to a curated prompt-rewriting SFT dataset with one command. The pipeline automatically:

  1. (Optional) Expands each seed prompt into more same-scenario prompts with topic dedup and lineage tracking.
  2. Rewrites every prompt through a three-step teacher agent chain: scene/language routing (plan), scene-specific rewriting (rewrite), and self-check correction (reflection).
  3. Scores each (original, rewritten) pair with a combined nine-metric LLM judge in a single call per row.
  4. Filters out rows that fail the score gates.
  5. Builds the final SFT dataset (system + user + assistant messages) with a per-language end-to-end student instruction.

Pipeline stages

Stage Required? Purpose
seed_anchored_expansion Optional Expand each seed into same-scenario prompts (topic dedup, source_seed_id/round lineage).
agentic_rewrite Required Teacher chain: plan (scene + language routing) -> rewrite (scene-specific system prompt) -> reflection (self-check).
pe_rewrite_eval Optional Combined nine-metric judge; one call per row appends all scores.
quality_filter Optional Drop rows below the score gates; optional per-scene top-k/top-ratio selection.
build_sft Required (last) Convert surviving rows into SFT messages with the student system prompt.

The last stage must always be build_sft.

Scene taxonomy and rewrite prompts

The plan step routes every prompt into one of 10 scenes: general, photographic_realism, artistic_illustration, design_layout, structured_diagram, ui_interface, brand_commercial_ad, narrative_panels, cultural_heritage_art, game_art_production.

Each (scene, language) pair has a scene-specific rewrite prompt at configs/prompts/pe_rewrite/rewrite_{scene}_{zh|en}.txt. Scenes without a dedicated file fall back to the general prompt of the same language (the two general files are required).

Shared iron laws (fidelity, information density, on-screen text expansion, quoting, language rules, self-check, output format) live in rewrite_common_{zh|en}.txt. When present, the common block of the matching language is automatically prepended to every scene prompt at load time; when absent, scene prompts are used as-is. Scene files only carry scene-specific additions, and clauses that intentionally deviate from a global iron law declare the override explicitly (marked with ⚠️).

Judge metrics and default filter gates

Seven 0-9 anchored metrics plus two boolean hard checks, all scored in one judge call per row:

Metric Type Default gate
intent_fidelity 0-9 >= 7
text_rendering_completeness 0-9 >= 7
usability 0-9 >= 7
detail_enrichment 0-9 >= 6
visual_concreteness 0-9 >= 6
compositional_coverage 0-9 >= 5
scene_alignment 0-9 >= 5
language_consistency bool must be true
no_conflict bool must be true

quality_filter applies these gates by default (override via min_scores). Adding keep_top_k / keep_top_ratio enables a second selection pass ranked by the average of the seven scores; it runs per scene by default (ceiling rounding, at least one row per scene) so no scene is evicted wholesale — set per_scene: false for global ranking.

Config schema

job_type: pe_rewrite_distill

backend:
  type: pai_token
  api_key: ${PAI_TOKEN_API_KEY}
  model_id: qwen3.7-plus         # default model for teacher steps

pipeline:
  - stage: seed_anchored_expansion   # omit when inputs are ready prompts
    config:
      rounds: 2
      generations_per_round: 5
    output_path: outputs/pe_expanded.jsonl
  - stage: agentic_rewrite
    config:
      reflection:
        model_id: qwen3.7-max        # stronger checker
      stream_output_path: outputs/pe_rewrite.stream.jsonl  # crash-safe sink
    output_path: outputs/pe_rewrite.jsonl
  - stage: pe_rewrite_eval
    config:
      model_id: qwen3.7-max          # judge separated from the teacher
      temperature: 0.0
    output_path: outputs/pe_scored.jsonl
  - stage: quality_filter
    config:
      per_scene: false               # hard gates only
    output_path: outputs/pe_filtered.jsonl
  - stage: build_sft

sft:
  system_prompt_zh_file: configs/prompts/pe_rewrite/student_system_zh.txt
  system_prompt_en_file: configs/prompts/pe_rewrite/student_system_en.txt

dataset:
  input_path: examples/seed_pe_prompts.jsonl
  output_path: outputs/pe_sft.jsonl

Reference configs:

  • From seeds: configs/pipeline/pe_rewrite_distill_from_seeds_pai_token.yaml (PAI-Token) and configs/pipeline/pe_rewrite_distill_from_seeds_pai_eas.yaml (PAI-EAS)
  • From ready prompts: configs/pipeline/pe_rewrite_distill_pai_token.yaml (PAI-Token) and configs/pipeline/pe_rewrite_distill_pai_eas.yaml (PAI-EAS)

The teacher (rewrite) and judge must run on different models sharing one backend endpoint (model_id overrides per stage) to avoid self-evaluation bias.

Output row schema

Intermediate rows keep every input field plus:

{"instruction": "...", "response": "...", "scene": "photographic_realism",
 "language": "zh", "agent_trace": {"plan": {}, "rewrite": {}, "reflection": {}, "durations": {}},
 "source_seed_id": "s1", "round": 0, "intent_fidelity": 8, "...": "..."}

The final SFT samples are {"messages": [system, user, assistant], "metadata": {...}}; judge scores and agent_trace are audit-only fields and never enter the SFT metadata, while expansion lineage does.

Standalone jobs

Every stage is also exposed as a standalone job_type for debugging or resuming from an intermediate JSONL:

Job LLM calls? Config example
seed_anchored_expansion Yes configs/rewrite/seed_anchored_expansion_pai_token.yaml
configs/rewrite/seed_anchored_expansion_pai_eas.yaml
agentic_rewrite Yes configs/rewrite/agentic_rewrite_pai_token.yaml
configs/rewrite/agentic_rewrite_pai_eas.yaml
pe_rewrite_eval Yes configs/rewrite/pe_rewrite_eval_pai_token.yaml
configs/rewrite/pe_rewrite_eval_pai_eas.yaml
pe_rewrite_filter No (local) configs/rewrite/pe_rewrite_filter.yaml
pe_rewrite_build_sft No (local) configs/rewrite/pe_rewrite_build_sft.yaml

The two local jobs need no backend section.