README.md

metadata

language:
  - en
license: cc-by-nc-4.0
pretty_name: Complexity Atlas Posttrain
task_categories:
  - text-generation
tags:
  - supervised-fine-tuning
  - post-training
  - card-based-generation
  - semantic-audit
configs:
  - config_name: chat
    data_files:
      - split: train
        path: data/sft-v13/chat/train-*.parquet
  - config_name: instruct
    data_files:
      - split: train
        path: data/sft-v13/instruct/train-*.parquet
      - split: validation
        path: data/sft-v13/instruct/validation.parquet
      - split: diagnostic
        path: data/sft-v13/instruct/diagnostic.parquet

Complexity Atlas Posttrain

Dataset release: v1.0.13

Complexity Atlas Posttrain is an original English post-training corpus built from linked semantic cards, intent-specific completion contracts, family-specific answer decks and auditable composition rules. The released prompts and responses use Complexity-authored material. No third-party conversation rows and no model-generated dialogue are included.

Model-facing SFT projections

The chat and instruct configurations expose the canonical v1.0.13 model-facing projection by interaction mode. Each row is stored once and belongs to exactly one configuration.

The projection removes exact prompt and response conflicts, selects compatible surface cards and applies the complexity-chat-v1 serialization contract. Historical source projections are excluded. Native 32k token shards remain available as derived training artifacts but are not Dataset Viewer inputs.

Split	Examples	Purpose
train	396,138	supervised fine-tuning
validation	28	separately authored held-out exchanges
diagnostic	672	deterministic family-coverage checks
total	396,838

Configuration	Train examples	Additional splits
chat	203,899	—
instruct	192,239	validation, diagnostic

The complete native-32k projection contains 94,815,656 serialized tokens, including 31,390,124 supervised assistant tokens. The training split contains 31,366,545 supervised tokens. Exact projected prompts and responses are unique.

The training data is stored in nine Parquet shards of at most 50,000 rows. Every file uses 5,000-row groups, Zstandard compression and a page index so the Hugging Face Dataset Viewer can read bounded previews without scanning a large monolithic file.

What changed in v1.0.13

The v1.0.13 projection adds an independent casual-conversation source instead of converting task instructions into artificial dialogue. It contributes 398 training conversations built from 420 original topic/context pairs, with four- and six-turn exchanges covering everyday observations, preferences, hobbies, small decisions and natural topic shifts. Source-pair groups cannot cross the train/validation boundary. The existing fourteen assistant families remain available alongside this additive source.

The casual-conversation source audit reports:

• 100% exact conversation and final-response uniqueness;
• zero source-pair overlap between train and validation;
• no surface hand or response structure above 5%;
• no four-word phrase above 5% of messages;
• 3.10% / 0% / 0% MiniLM semantic-neighbor ratios for prompts, responses and complete conversations at cosine 0.98;
• 0.5% for all three views with Mixedbread at the same threshold.

Tokenized native 32k shards

tokenized/32k-v13/ contains the exact v1.0.13 projection for the native 32k tokenizer used by the 306.5M checkpoint:

• little-endian uint32 input IDs;
• aligned little-endian int32 labels;
• -100 for system and user positions;
• assistant-only causal loss and an EOS target;
• per-example offsets in examples.jsonl and sft.idx.json;
• the required complexity-chat-v1 template in chat_template.json.

Parquet remains the canonical readable form. The 28-row validation split is independently authored. The 672-row diagnostic split is deterministic and must not be presented as independent human evaluation.

Limitations

• The corpus is experimental and English-only.
• The held-out evaluation set is small and is not a comprehensive benchmark.
• Automated structural checks do not replace downstream model evaluation.
• Casual dialogue remains a small additive source rather than the majority of the full projection.
• Embedding similarity is a diagnostic and not proof of conversational correctness or independence.

Intended use

This dataset is intended for experimental supervised post-training of small English language models. Evaluate general-language regression, repetition, held-out behavior and safety before deployment.

Provenance and license

The original cards, schemas, compositions, curation and released dataset artifacts are licensed under CC BY-NC 4.0.

Attribution: Complexity — Complexity Atlas Posttrain

The Apache-2.0 builder and audit code is maintained at https://github.com/Complexity-ML/complexity-card-corpus.