EasyDistill 2 JSONL Data Formats

This document describes the JSONL schemas used as inputs and outputs across EasyDistill 2 pipelines and standalone jobs. All JSONL files use UTF-8 encoding and contain one valid JSON object per line.

Common input formats

Seed instructions

Used by instruction-distillation jobs and pipelines.

{"instruction": "What is the capital of France?"}
{"instruction": "Explain quantum computing in one sentence.", "system": "You are a concise tutor."}

Fields:

instruction (string, required): the user prompt.
system (string, optional): per-row system prompt. Falls back to the config-level system_prompt.
id (string/integer, optional): row identifier; auto-generated if omitted.

Seed problems for CoT

Used by cot_distill and advanced_cot_distill.

{"problem": "What is the sum of the first 10 positive integers?"}
{"instruction": "What is 2+2?"}

The problem field is configurable via dataset.problem_key (default problem, fallback instruction).

Problem / answer pairs for CoT rewriting

Used by cot_long2short and cot_short2long.

{"instruction": "What is 2+2?", "response": "<|begin_of_thought|>...<|end_of_thought|><|begin_of_solution|>4<|end_of_solution|>"}

Keys are configurable via dataset.problem_key (default instruction) and dataset.answer_key (default response). Common fallbacks (problem, answer, output) are also accepted.

Multi-modal inputs

Used by mm_instruct_distill and mm_cot_distill.

{"id": "mm_0", "instruction": "Describe what you see.", "images": ["examples/mm_sample_image.png"]}
{"id": "mm_1", "instruction": "What color is the main object?", "images": ["https://example.com/img.png"]}

Fields:

instruction (string, required): the text prompt.
images (list of strings, required): image references. Each item may be a local path, file:// URI, http(s):// URL, or base64 data URL such as data:image/png;base64,....
id (optional): row identifier.

Evaluation inputs

Used by instruct_eval, cot_eval, mm_instruct_eval, and mm_cot_eval.

Plain format:

{"instruction": "What is the capital of France?", "output": "Paris"}

SFT messages format (auto-converted):

{"messages": [{"role": "user", "content": "What is the capital of France?"}, {"role": "assistant", "content": "Paris"}]}

For multi-modal evaluation, images may also be present.

Raw text for response extraction

Used by instruction_response_extraction.

{"text": "User: What is 2+2?\nAssistant: 2+2 equals 4."}

DPO seed prompts

For dpo_instruct_*:

{"instruction": "Explain knowledge distillation in one paragraph."}

For dpo_cot_*:

{"problem": "What is the sum of the first 10 positive integers?", "answer": "55"}

Fields are configurable via instruction_key / answer_key.

Common output formats

SFT messages

Produced by any job that ends with build_sft or by standalone distillation jobs such as instruct_distill, cot_distill, mm_instruct_distill, and mm_cot_distill.

Text-only SFT:

{
  "messages": [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "What is 2+2?"},
    {"role": "assistant", "content": "4"}
  ],
  "metadata": {
    "source": "teacher_model",
    "model": "Qwen2.5-3B-Instruct",
    "request_id": "1",
    "backend": "pai_eas",
    "usage": {"completion_tokens": 1, "prompt_tokens": 31, "total_tokens": 32}
  }
}

Multi-modal SFT:

{
  "messages": [
    {"role": "system", "content": "You are a helpful visual assistant."},
    {"role": "user", "content": [
      {"type": "image_url", "image_url": {"url": "data:image/png;base64,..."}},
      {"type": "text", "text": "Describe what you see."}
    ]},
    {"role": "assistant", "content": "The image shows a solid red square."}
  ],
  "metadata": {
    "source": "teacher_model",
    "model": "Qwen2.5-VL-3B-Instruct",
    "request_id": "mm_gen_0",
    "backend": "pai_eas",
    "instruction": "Describe what you see.",
    "images": ["examples/mm_sample_image.png"],
    "usage": {"completion_tokens": 235, "prompt_tokens": 103, "total_tokens": 338}
  }
}

Fields:

messages (list, required): OpenAI/ShareGPT message objects. Multi-modal user messages contain a list of image_url and text content items.
metadata (dict, optional): provenance information such as source, model, request_id, backend, usage, and original instruction / images.

Evaluation outputs

Produced by instruct_eval and cot_eval.

{"id": "0", "instruction": "What is the capital of France?", "output": "Paris", "informativeness": 2, "helpfulness": 7, "generalization": 1, "correctness": true}

{"id": "0", "instruction": "What is the sum of the first 10 positive integers?", "output": "...", "reasoning_verbosity": 5, "cognitive_difficulty": 5, "logical_correctness": true}

The original row fields are preserved and the requested metrics are appended.

Instruction expansion output

Produced by instruction_expansion.

{"instruction": "Write a new instruction similar to the examples but different in content."}

Instruction refinement output

Produced by instruction_refinement.

{"instruction": "Rewrite the input instruction to be clearer and more specific."}

Instruction balancing output

Produced by instruction_balance.

{"instruction": "What is 2+2?", "category": "Math"}

The original fields are preserved and a category field is added.

Generated response rows (before SFT conversion)

Produced by the generate stage inside pipelines.

{"instruction": "What is 2+2?", "output": "4"}

Quality-filtered rows

Produced by the quality_filter stage inside pipelines. The format is the same as the evaluated rows, with only rows that pass the thresholds retained.

{"instruction": "What is 2+2?", "output": "4", "correctness": true, "helpfulness": 7}

CoT RV/CD scored rows

Produced by the cot_rvcd_score stage.

{
  "instruction": "What is the sum of the first 10 positive integers?",
  "response": "...",
  "reasoning_verbosity": 5,
  "cognitive_difficulty": 4,
  "logical_correctness": true
}

CoT RV/CD mixed rows

Produced by the cot_mix_by_rv_cd stage.

{
  "instruction": "...",
  "response": "...",
  "reasoning_verbosity": 2,
  "cognitive_difficulty": 2,
  "logical_correctness": true,
  "cd_bin": 0,
  "rv_target": 2.0
}

Agent distillation formats

Used by the agent_distill pipeline.

Agent seed personas

Input to agent_task_synthesis.

{"id": "persona_001", "background": "An Afrikaans music fan who wants to organize local events."}

Fields:

id (string/integer, optional): row identifier.
background or persona (string, required): persona or background description.

`agent_task_synthesis` output

{
  "id": "persona_001",
  "background": "An Afrikaans music fan...",
  "task": "Plan a local concert for Afrikaans music.",
  "tools": [{"name": "search_venues", "description": "Search venues"}],
  "workflow": "1. Find venues 2. Book artists 3. Promote",
  "restriction": "Stay within the stated budget.",
  "initial_toolset_create": "<task>...</task><tools>...</tools>..."
}

`agent_fuzzy_task` output

{
  "id": "persona_001",
  "fuzzy_task": "Help organize a small concert on a limited budget.",
  "task_background": "The user is an Afrikaans music fan with no event-planning experience...",
  "raw_fuzzy_task": "<task>...</task><background>...</background>"
}

`agent_tool_check` output

{
  "id": "persona_001",
  "checked_tools": [{"name": "search_venues", "description": "Search venues"}],
  "raw_tool_check": "<tools>...</tools>"
}

`agent_trajectory` output

One row per rollout.

{
  "id": "persona_001",
  "solution_id": "persona_001_solution_1.json",
  "fuzzy_task": "Help organize a small concert on a limited budget.",
  "task_background": "...",
  "restriction": "Stay within budget.",
  "checked_tools": [{"name": "search_venues"}],
  "trajectory": [
    {"role": "system", "content": "You are a helpful agent."},
    {"role": "user", "content": "Help organize a small concert on a limited budget."},
    {"role": "assistant", "content": "I will search for venues.<tool_call>{...}</tool_call>"},
    {"role": "user", "content": "<tool_response>Found 3 venues.</tool_response>"},
    {"role": "assistant", "content": "<answer>Book the Community Hall.</answer>"}
  ],
  "tool_call_history": ["Query:\n...\nResponse:\n..."],
  "task_finished": "Terminated"
}

`agent_rubrics` output

One row per task, grouping trajectories and selecting the best solution.

{
  "id": "persona_001",
  "fuzzy_task": "Help organize a small concert on a limited budget.",
  "best_solution_id": "persona_001_solution_1.json",
  "alignment_check": "The trajectories align with the task.",
  "rubrics": "1. Correctness 2. Efficiency",
  "final": "Solution 1 is best.",
  "trajectories": [...]
}

`build_sft` output for agent distillation

Standard SFT messages format. The metadata block includes task_id, solution_id, task_finished, task, fuzzy_task, restriction, and workflow.

`build_preference_dataset` output for agent distillation

{
  "prompt": "Help organize a small concert on a limited budget.",
  "chosen": "[{...best trajectory messages...}]",
  "rejected": "[{...worst trajectory messages...}]",
  "system": "You are a helpful assistant..."
}

DPO intermediate formats

`generate_candidates` output

{
  "id": "1",
  "instruction": "Explain knowledge distillation in one paragraph.",
  "candidates": ["...", "..."],
  "candidate_results": [
    {"request": {...}, "response": "...", "model": "...", "usage": {...}, "metadata": {...}},
    {"request": {...}, "response": "...", "model": "...", "usage": {...}, "metadata": {...}}
  ]
}

For CoT DPO, the row contains problem and answer instead of instruction.

`score_candidates` output

Same as candidates output with candidate_scores and, for the CoT scorer, candidate_correctness added.

{
  "id": "1",
  "instruction": "Explain knowledge distillation in one paragraph.",
  "candidates": ["...", "..."],
  "candidate_scores": [4.0, 4.0]
}

`build_preference_pairs` output

{
  "id": "1",
  "instruction": "Explain knowledge distillation in one paragraph.",
  "system": null,
  "chosen": "...",
  "rejected": "...",
  "chosen_score": 4.0,
  "rejected_score": 4.0,
  "answer": null
}

For CoT DPO, instruction is replaced by problem and answer contains the reference answer.

`build_preference_dataset` output

llama_factory_alpaca:

{"instruction": "...", "input": "", "chosen": "...", "rejected": "..."}

llama_factory_sharegpt:

{
  "conversations": [
    {"from": "human", "value": "..."},
    {"from": "gpt", "value": "..."}
  ],
  "chosen": {"from": "gpt", "value": "..."},
  "rejected": {"from": "gpt", "value": "..."}
}

openai_messages:

{
  "prompt": [{"role": "user", "content": "..."}],
  "chosen": [{"role": "assistant", "content": "..."}],
  "rejected": [{"role": "assistant", "content": "..."}]
}

Multi-modal CoT rewrite input/output

mm_cot_long2short and mm_cot_short2long accept both raw rows and SFT message rows.

Raw input:

{"instruction": "Look at the image and determine the dominant color.", "images": ["examples/mm_sample_image.png"], "response": "..."}

SFT message input (auto-converted, images read from metadata.images):

{
  "messages": [
    {"role": "user", "content": [{"type": "image_url", ...}, {"type": "text", ...}]},
    {"role": "assistant", "content": "..."}
  ],
  "metadata": {"images": ["examples/mm_sample_image.png"]}
}

mm_cot_long2short output includes response (simplified), original_response, original_tokens, simplified_tokens, and compression_ratio.

mm_cot_short2long output includes response (extended), original_response, original_tokens, extended_tokens, expansion_ratio, and step_count.

T2I distillation formats

For T2I distillation input/output schemas and stage-by-stage JSONL formats, see t2i_distillation.md for an overview and t2i_distillation_implementation.md for the complete data-flow schemas.

PE rewrite distillation formats

Seed PE prompts

Input for pe_rewrite_distill and seed_anchored_expansion (see examples/seed_pe_prompts.jsonl). id is optional and used for expansion lineage; the key is configurable via dataset.instruction_key (default instruction):

{"id": "pe_seed_001", "instruction": "画一张水循环的科普信息图，包含蒸发、凝结、降水几个环节，中文标注，图标简洁一点"}

`seed_anchored_expansion` output

One row per generated prompt, with lineage back to the source seed and the round-level dedup topic:

{"instruction": "画一张光合作用原理的科普长图...", "source_seed_id": "pe_seed_001", "round": 0, "topic": "光合作用原理图解"}

`agentic_rewrite` output

Adds the final rewritten prompt (response), the plan routing result (scene / language) and an agent_trace audit object; extra input fields (e.g. expansion lineage) pass through unchanged:

{"instruction": "画一张水循环的科普信息图...", "response": "一张竖版科普信息图，主标题\"水循环\"位于顶部...", "scene": "structured_diagram", "language": "zh", "agent_trace": {"plan": {"status": "ok", "raw": "..."}, "rewrite": {"status": "ok", "draft": "..."}, "reflection": {"status": "ok", "changed": false, "notes": "", "raw": "..."}, "durations": {"plan": 1.2, "rewrite": 8.5, "reflection": 3.1}}, "source_seed_id": "pe_seed_001", "round": 0, "topic": "..."}

`pe_rewrite_eval` output

Adds seven 0-9 integer metrics and two boolean hard checks to every row (unparseable metrics come back as null):

{"instruction": "...", "response": "...", "scene": "structured_diagram", "language": "zh", "intent_fidelity": 8, "text_rendering_completeness": 9, "detail_enrichment": 8, "visual_concreteness": 8, "compositional_coverage": 7, "scene_alignment": 8, "usability": 9, "language_consistency": true, "no_conflict": true, "agent_trace": {"...": "..."}, "source_seed_id": "pe_seed_001", "round": 0}

The pe_rewrite_filter stage keeps rows passing the score gates (plus the optional per-scene top selection) without changing the row schema.

`pe_rewrite_build_sft` output

SFT rows whose system message is the per-language student rewrite instruction. Judge scores and agent_trace are audit-only and never enter metadata, while scene routing and expansion lineage are carried over:

{
  "messages": [
    {"role": "system", "content": "你是文生图 prompt 改写专家..."},
    {"role": "user", "content": "画一张水循环的科普信息图..."},
    {"role": "assistant", "content": "一张竖版科普信息图，主标题\"水循环\"位于顶部..."}
  ],
  "metadata": {"source": "teacher_model", "model": "pipeline", "request_id": "0", "scene": "structured_diagram", "language": "zh", "source_seed_id": "pe_seed_001", "round": 0, "topic": "..."}
}

Uh oh!

FilesExpand file tree

data_formats.md

Latest commit

History

data_formats.md

File metadata and controls

EasyDistill 2 JSONL Data Formats

Common input formats

Seed instructions

Seed problems for CoT

Problem / answer pairs for CoT rewriting

Multi-modal inputs

Evaluation inputs

Raw text for response extraction

DPO seed prompts

Common output formats

SFT messages

Evaluation outputs

Instruction expansion output

Instruction refinement output

Instruction balancing output

Generated response rows (before SFT conversion)

Quality-filtered rows

CoT RV/CD scored rows

CoT RV/CD mixed rows

Agent distillation formats

Agent seed personas

agent_task_synthesis output

agent_fuzzy_task output

agent_tool_check output

agent_trajectory output

agent_rubrics output

build_sft output for agent distillation

build_preference_dataset output for agent distillation

DPO intermediate formats

generate_candidates output

score_candidates output

build_preference_pairs output

build_preference_dataset output

Multi-modal CoT rewrite input/output

T2I distillation formats

PE rewrite distillation formats

Seed PE prompts

seed_anchored_expansion output

agentic_rewrite output

pe_rewrite_eval output

pe_rewrite_build_sft output

`agent_task_synthesis` output

`agent_fuzzy_task` output

`agent_tool_check` output

`agent_trajectory` output

`agent_rubrics` output

`build_sft` output for agent distillation

`build_preference_dataset` output for agent distillation

`generate_candidates` output

`score_candidates` output

`build_preference_pairs` output

`build_preference_dataset` output

`seed_anchored_expansion` output

`agentic_rewrite` output

`pe_rewrite_eval` output

`pe_rewrite_build_sft` output