Persona MCQ Configuration Reference#

Configuration fields, output artifacts, and named errors for sdg/persona_mcq. For the procedure, refer to Generate Persona-Grounded MCQ Data for SFT.

Command#

uv run nemotron steps run sdg/persona_mcq -c {default|tiny} [-d] [--run PROFILE | --batch PROFILE] [KEY=VALUE ...]

Overrides use OmegaConf dotlist syntax. List values are passed in brackets and quoted for the shell, for example 'pipeline.stages=[answers,build_sft,sample]'. Shipped environment profiles are <backend>_sdg_persona_mcq and <backend>_sdg_persona_mcq_tiny for lepton, slurm, and dgxcloud.

Shipped Configuration#

pipeline:
  experiment_name: persona-mcq
  output_root: ${oc.env:NEMOTRON_RUN_DIR,./outputs}/persona_mcq
  stages: [all]
  seed: 42
  resume: true
  overwrite: false

languages:
  english:
    display_name: English
    locale: en_IN
    answer_label: Answer
    script_pattern: "[A-Za-z]"
    question_script_fraction: {min: 0.50, max: 1.0}
    answer_script_fraction: {min: 0.85, max: 1.0}
    strip_latin_glosses: false
  hindi:
    display_name: Hindi
    locale: hi_Deva_IN
    answer_label: उत्तर
    answer_label_aliases: [उत्तर है]
    script_pattern: "[ऀ-ॿ]"
    question_script_fraction: {min: 0.50, max: 1.0}
    answer_script_fraction: {min: 0.70, max: 1.0}
    strip_latin_glosses: true
  malayalam:
    display_name: Malayalam
    # Data Designer has no managed Malayalam persona locale; use Indian
    # personas as grounding while the model authors and answers in Malayalam.
    locale: en_IN
    answer_label: ഉത്തരം
    script_pattern: "[ഀ-ൿ]"
    question_script_fraction: {min: 0.50, max: 1.0}
    answer_script_fraction: {min: 0.70, max: 1.0}
    strip_latin_glosses: true

question_models: [qwen, oss, gemma]
answer_models: [qwen, oss, gemma]

models:
  qwen:
    model: Qwen/Qwen3.5-122B-A10B
    endpoint: ${oc.env:QWEN_API_BASE}
    api_key_env: NVIDIA_API_KEY
    skip_health_check: true
    question:
      inference:
        temperature: 0.7
        top_p: 0.8
        max_tokens: 4096
        max_parallel_requests: 256
        timeout: 300
        extra_body:
          top_k: 20
          presence_penalty: 1.5
          chat_template_kwargs: {enable_thinking: false}
    answer:
      temperature: 1.0
      top_p: 0.95
      max_tokens: 16384
      extra_body:
        top_k: 20
        presence_penalty: 1.5
        chat_template_kwargs: {enable_thinking: true}
  oss:
    model: openai/gpt-oss-120b
    endpoint: ${oc.env:OSS_API_BASE}
    api_key_env: NVIDIA_API_KEY
    skip_health_check: true
    question:
      inference:
        temperature: 1.0
        top_p: 1.0
        max_tokens: 4096
        max_parallel_requests: 256
        timeout: 300
        extra_body: {reasoning_effort: low}
    answer:
      temperature: 1.0
      top_p: 1.0
      max_tokens: 16384
      extra_body: {reasoning_effort: high}
  gemma:
    model: google/gemma-4-31B-it
    endpoint: ${oc.env:GEMMA_API_BASE}
    api_key_env: NVIDIA_API_KEY
    skip_health_check: true
    question:
      inference:
        temperature: 1.0
        top_p: 0.95
        max_tokens: 4096
        max_parallel_requests: 256
        timeout: 300
        extra_body:
          top_k: 64
          chat_template_kwargs: {enable_thinking: false}
    answer:
      temperature: 1.0
      top_p: 0.95
      max_tokens: 16384
      extra_body:
        top_k: 64
        chat_template_kwargs: {enable_thinking: true}

question_generation:
  num_records: 100000
  max_parallel_requests: 256
  buffer_size: 500

lexical_dedup:
  threshold: 0.80
  shingle_size: 4
  permutations: 128
  bands: 16
  seed: 42

semantic_dedup:
  model: intfloat/multilingual-e5-large
  device: cuda
  batch_size: 256
  threshold: 0.965
  method: greedy
  seed: 13
  chunk_size: 4096

answer_generation:
  max_parallel_requests: 256
  timeout: 600
  max_retries: 5

sft:
  response_teachers: [qwen, oss, gemma]
  agreement: unanimous
  reasoning:
    display_name: English
    script_pattern: "[A-Za-z]"
    script_fraction: {min: 0.85, max: 1.0}

sampling:
  per_language: 50000
  seed: 42
  reasoning_off_fraction: 0.10
  answer_variant: full

Key differences in tiny.yaml include pipeline.experiment_name (persona-mcq-smoke), question_models ([oss]), reduced record, buffer, parallelism, token, MinHash, and semantic-dedup sizes, semantic_dedup.device: cpu, answer_generation.max_retries: 3, and sampling.per_language: 1. Refer to the complete configuration when exact smoke-test values matter:

pipeline:
  experiment_name: persona-mcq-smoke
  output_root: ${oc.env:NEMOTRON_RUN_DIR,./outputs}/persona_mcq
  stages: [all]
  seed: 42
  resume: true
  overwrite: false

languages:
  english:
    display_name: English
    locale: en_IN
    answer_label: Answer
    script_pattern: "[A-Za-z]"
    question_script_fraction: {min: 0.50, max: 1.0}
    answer_script_fraction: {min: 0.85, max: 1.0}
    strip_latin_glosses: false
  hindi:
    display_name: Hindi
    locale: hi_Deva_IN
    answer_label: उत्तर
    answer_label_aliases: [उत्तर है]
    script_pattern: "[ऀ-ॿ]"
    question_script_fraction: {min: 0.50, max: 1.0}
    answer_script_fraction: {min: 0.70, max: 1.0}
    strip_latin_glosses: true
  malayalam:
    display_name: Malayalam
    locale: en_IN
    answer_label: ഉത്തരം
    script_pattern: "[ഀ-ൿ]"
    question_script_fraction: {min: 0.50, max: 1.0}
    answer_script_fraction: {min: 0.70, max: 1.0}
    strip_latin_glosses: true

question_models: [oss]
answer_models: [qwen, oss, gemma]

models:
  qwen:
    model: Qwen/Qwen3.5-122B-A10B
    endpoint: ${oc.env:QWEN_API_BASE}
    api_key_env: NVIDIA_API_KEY
    skip_health_check: true
    question: {inference: {temperature: 0.7, top_p: 0.8, max_tokens: 2048, max_parallel_requests: 4, timeout: 300}}
    answer: {temperature: 1.0, top_p: 0.95, max_tokens: 8192, extra_body: {chat_template_kwargs: {enable_thinking: true}}}
  oss:
    model: openai/gpt-oss-120b
    endpoint: ${oc.env:OSS_API_BASE}
    api_key_env: NVIDIA_API_KEY
    skip_health_check: true
    question: {inference: {temperature: 1.0, top_p: 1.0, max_tokens: 2048, max_parallel_requests: 4, timeout: 300, extra_body: {reasoning_effort: low}}}
    answer: {temperature: 1.0, top_p: 1.0, max_tokens: 8192, extra_body: {reasoning_effort: high}}
  gemma:
    model: google/gemma-4-31B-it
    endpoint: ${oc.env:GEMMA_API_BASE}
    api_key_env: NVIDIA_API_KEY
    skip_health_check: true
    question: {inference: {temperature: 1.0, top_p: 0.95, max_tokens: 2048, max_parallel_requests: 4, timeout: 300}}
    answer: {temperature: 1.0, top_p: 0.95, max_tokens: 8192, extra_body: {chat_template_kwargs: {enable_thinking: true}}}

question_generation: {num_records: 8, max_parallel_requests: 4, buffer_size: 8}
lexical_dedup: {threshold: 0.80, shingle_size: 4, permutations: 32, bands: 4, seed: 42}
semantic_dedup: {model: intfloat/multilingual-e5-large, device: cpu, batch_size: 8, threshold: 0.965, method: greedy, seed: 13, chunk_size: 64}
answer_generation: {max_parallel_requests: 4, timeout: 600, max_retries: 3}
sft:
  response_teachers: [qwen, oss, gemma]
  agreement: unanimous
  reasoning:
    display_name: English
    script_pattern: "[A-Za-z]"
    script_fraction: {min: 0.85, max: 1.0}
sampling: {per_language: 1, seed: 42, reasoning_off_fraction: 0.10, answer_variant: full}

Fields#

`pipeline`#

Field	Default	Description
`experiment_name`	`persona-mcq`	Required. Artifact directory name under `output_root`; also identifies the configuration for resume checks.
`output_root`	`${NEMOTRON_RUN_DIR:-./outputs}/persona_mcq`	Parent directory for experiments.
`stages`	`[all]`	`[all]` or an ordered subset of `personas`, `questions`, `lexical_dedup`, `semantic_dedup`, `answer_seed`, `answers`, `build_sft`, `sample`. `all` cannot be combined with named stages.
`seed`	`42`	Pipeline seed used by question authoring. Answer option shuffles are content-deterministic from `query_id` and do not change with this value.
`resume`	`true`	Skip completed question outputs and answered rows; reload the existing `summary.json`.
`overwrite`	`false`	Delete the experiment directory before running. Required to reuse an experiment name with a different configuration.

`languages.<key>`#

Every entry is independent; the key names carry no language-specific behavior.

Field	Required	Description
`display_name`	yes	Language name inserted into authoring and answering prompts.
`locale`	yes	Data Designer persona locale, for example `en_IN` or `hi_Deva_IN`.
`answer_label`	yes	Localized final-answer marker; the assistant’s last line must be `<answer_label>: <LETTER>`.
`answer_label_aliases`	no	Additional accepted markers when parsing teacher answers.
`script_pattern`	yes	Single-character regular expression for the target script. Must not match the empty string.
`question_script_fraction`	yes	`{min, max}` inclusive range for the fraction of alphabetic characters in a question that match `script_pattern`.
`answer_script_fraction`	yes	The same range applied to the exported assistant answer.
`strip_latin_glosses`	no (`false`)	Remove parenthetical Latin-script glosses from questions and options before deduplication.

`question_models`, `answer_models`, and `models.<alias>`#

question_models lists the aliases whose questions are pooled before deduplication. answer_models lists the teacher panel; at least three are required. Every alias must be defined under models.

Field	Required	Description
`model`	yes	Model identifier sent to the endpoint.
`endpoint`	yes	OpenAI-compatible base URL; the shipped configurations resolve `${oc.env:*_API_BASE}`.
`api_key_env`	yes	Name of the environment variable holding the bearer token. The value is never written to run metadata.
`provider_type`	no (`openai`)	Data Designer provider type for question authoring.
`skip_health_check`	no (`false`)	Skip the Data Designer endpoint health check.
`extra_headers`	no	Additional HTTP headers for question authoring.
`question.inference`	no	Data Designer `ChatCompletionInferenceParams` for authoring: `temperature`, `top_p`, `max_tokens`, `max_parallel_requests`, `timeout`, `extra_body`.
`answer`	no	Answer request parameters: `temperature` (`1.0`), `top_p` (`1.0`), `max_tokens` (`16384`), and `extra_body` merged into the request body.

`question_generation`#

Field	Default	Description
`num_records`	`100000`	Questions requested per question model and language.
`max_parallel_requests`	`256`	Data Designer non-inference worker parallelism.
`buffer_size`	`500`	Data Designer run buffer size.

`lexical_dedup`#

Field	Default	Description
`threshold`	`0.80`	Jaccard similarity of question shingles at or above which a question is a near-duplicate.
`shingle_size`	`4`	Shingle length.
`permutations`	`128`	MinHash permutations. Must be divisible by `bands`.
`bands`	`16`	Locality-sensitive hashing bands.
`seed`	`42`	MinHash seed.

`semantic_dedup`#

Field	Default	Description
`model`	`intfloat/multilingual-e5-large`	Sentence Transformers embedding model. Questions are embedded with the `query:` prefix.
`device`	`cuda`	Embedding device. `tiny.yaml` uses `cpu`.
`batch_size`	`256`	Embedding batch size.
`threshold`	`0.965`	Cosine-similarity threshold for a near-duplicate pair; must be between 0 and 1.
`method`	`greedy`	`greedy` visits records in a seeded random order, keeps each unvisited record, and removes its similar neighbors; `components` keeps one randomly chosen record per connected component of similar pairs.
`seed`	`13`	Seed for the keep order.
`chunk_size`	`4096`	Rows per similarity block.

`answer_generation`#

Field	Default	Description
`max_parallel_requests`	`256`	Concurrent answer requests per teacher and language.
`timeout`	`600`	HTTP timeout in seconds.
`max_retries`	`5`	Attempts per row; HTTP 429 and 5xx responses are retried with exponential back-off capped at 30 seconds.

`sft`#

Field	Default	Description
`response_teachers`	`[qwen, oss, gemma]`	Teachers whose answers are exported. Each must also appear in `answer_models`.
`agreement`	`unanimous`	`unanimous`: every parsed letter must match. `majority`: strictly more than half of the panel must match. The exporting teacher must match the vote.
`reasoning.display_name`	`English`	Language requested for the reasoning.
`reasoning.script_pattern`	`[A-Za-z]`	Script matcher for the reasoning.
`reasoning.script_fraction`	`{min: 0.85, max: 1.0}`	Inclusive script-fraction range for the retained reasoning.

`sampling`#

Field	Default	Description
`per_language`	`50000`	Records sampled per language from the questions that every response teacher exported.
`seed`	`42`	Sampling and shuffle seed.
`reasoning_off_fraction`	`0.10`	Fraction of sampled records per language that omit `reasoning_content`; the selection is deterministic in `seed` and `query_id`.
`answer_variant`	`full`	`full` keeps the teacher’s answer text; `stripped` replaces it with `<answer_label>: <letter>`.

Output Artifacts#

All paths are relative to <output_root>/<experiment_name>/.

Path	Content
`run.json`	Redacted configuration, including resolved endpoint URLs but not API-key values, plus `config_hash`, `data_designer_version`, and `nemotron_commit`.
`summary.json`	Per-stage statistics: cached and downloaded locales, question counts, deduplication drop reasons, answer counts, `build_sft` rejection reasons, and sample views. Rewritten after every stage.
`questions/<model>/<language>/records.jsonl`	Authored question records.
`questions/<model>/<language>/artifacts/`	Data Designer run artifacts for question authoring.
`lexical/<language>.jsonl`	Pooled, lexically deduplicated questions with `query_id`, `question`, `choices`, and provenance `metadata`.
`semantic/<language>.jsonl`	Semantically deduplicated questions.
`answer_seed/<language>.jsonl`	Shuffled options with `original_choices` and `shuffle_permutation`.
`answers/<model>/<language>/answers.jsonl`	Teacher responses with `answer`, `reasoning`, `parsed_letter`, `finish_reason`, and `completion_tokens`. Append-safe on resume.
`answers/<model>/<language>/failures.jsonl`	Rows that failed after `max_retries`; retried on resume.
`sft/<teacher>/<language>.jsonl`	Quality-gated `{messages, metadata}` records.
`training/<teacher>/train.jsonl`, `blend.json`	Aligned samples across all configured languages and the corresponding SFT-packing blend manifest.
`training/<teacher>/english_<language>/train.jsonl`, `blend.json`	Equal-count English/target views, written when a language named `english` is configured.

Training Record#

{
  "messages": [
    {"role": "system", "content": ""},
    {"role": "user", "content": "<answer instruction>\n\n<question>\n\nA) ...\nB) ...\nC) ...\nD) ..."},
    {"role": "assistant", "reasoning_content": "<English reasoning>", "content": "<answer text ending in answer_label: LETTER>"}
  ],
  "metadata": {
    "query_id": "...",
    "language": "hindi",
    "answer_label": "उत्तर",
    "reasoning_language": "English",
    "response_model": "gemma",
    "agreement": "unanimous",
    "voted_letter": "B",
    "n_valid_votes": 3,
    "difficulty": "...",
    "topic": "...",
    "facet": "...",
    "region": "...",
    "gen_model": "oss",
    "persona_uuid": "...",
    "reasoning_mode": "on"
  }
}

Records with reasoning_mode: "off" omit reasoning_content.

Persona Asset Cache#

The personas stage looks for each configured languages.<key>.locale at:

<managed-assets>/datasets/<locale>.parquet

DATA_DESIGNER_MANAGED_ASSETS_PATH sets <managed-assets> directly. If it is unset, the path defaults to ${DATA_DESIGNER_HOME:-~/.data-designer}/managed-assets. To populate the cache without giving the pipeline an NGC key, set the same paths in the downloading shell and download every configured locale before running:

$ export DATA_DESIGNER_HOME="<shared-cache-root>"
$ export DATA_DESIGNER_MANAGED_ASSETS_PATH="$DATA_DESIGNER_HOME/managed-assets"
$ uv run data-designer download personas --locale en_US

Repeat the download for each configured locale, then make the cache visible to the pipeline job at the same path. The shipped *_sdg_persona_mcq profiles forward both variables.

Errors#

Error	Cause	Recovery
`experiment_config_mismatch`	The experiment directory was created with a different configuration (`config_hash` differs).	Choose a new `pipeline.experiment_name`, or set `pipeline.overwrite=true` intentionally. Never combine artifacts from incompatible configurations.
`persona_assets_missing`	A configured locale has no cached persona asset and `NGC_API_KEY` is unset or the NGC CLI is not on `PATH`.	Export `NGC_API_KEY` and install the NGC CLI, or pre-download every locale as described in Persona Asset Cache.
`teacher_intersection_too_small`	Fewer questions survived every response teacher’s gates than `sampling.per_language` requests.	Inspect the `build_sft` rejection reasons in `summary.json`; reduce `sampling.per_language`, relax `sft.agreement`, or resume the `answers` stage to retry failed rows.

Configuration validation also rejects fewer than three answer_models, response teachers absent from answer_models, unknown stage names, invalid script_pattern values, and script-fraction ranges outside 0 <= min <= max <= 1. The lexical_dedup stage separately rejects permutations values that are not divisible by bands when that stage runs; check the pair before a full run because initial configuration validation does not inspect it.

Persona MCQ Configuration Reference#

Command#

Shipped Configuration#

Fields#

pipeline#

languages.<key>#

question_models, answer_models, and models.<alias>#

question_generation#

lexical_dedup#

semantic_dedup#

answer_generation#

sft#

sampling#