Pacific-i64's picture
Publish capability-audited Card Corpus V2 229k
456b0b5 verified
|
Raw
History Blame Contribute Delete
5.14 kB
metadata
license: cc-by-nc-4.0
language:
  - en
task_categories:
  - text-generation
  - question-answering
tags:
  - sft
  - conversational
  - reasoning
  - variableby2d
  - assistant-only-loss
pretty_name: Complexity Atlas Posttrain  Card Corpus V2
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet
      - split: validation
        path: data/validation-*.parquet
      - split: test
        path: data/test-*.parquet

Complexity Atlas Posttrain — Card Corpus V2

An English supervised fine-tuning corpus generated from authored semantic frames, role-separated prompt/answer/thinking plans, compatibility graphs, and VariableBy2D reservoirs. All 15 task families, including natural dialogue, belong to one audited corpus and one tokenizer-compatible training view.

Release

Split Examples
Train 224,654
Validation 2,478
Test 1,894
Total 229,026

The generator renders every registered valid scenario combination. No global sampling quota, per-family truncation, or 400K cap is applied.

Family Examples
brainstorming_creativity 384
casual_conversation 52,794
context_clarification 9,216
conversation_empathy 2,048
critique_revision 64
explanation_learning 36,040
extraction_classification 3,456
grounded_qa 4,608
planning_comparison 384
practical_action 384
reasoning_verification 108,000
safety_uncertainty 3,072
summarization_synthesis 4,096
troubleshooting 384
writing_transformation 4,096

Behavioral coverage

The release contract requires learnable support—not merely one public anchor— for the behaviors used during model promotion.

Capability Training examples Domain count
Direct safety 3,077 16
Small arithmetic 16,101 4
Summarization 4,096 8
Writing transformation 4,096 8
Multi-constraint following 4,000 4
Concept definitions 1,000 1
General facts 1,024 1
Reflective conversation 2,048 10
Neutral greetings 1,100 1

The casual family also contains 11,000 history-dependent multi-turn examples. Earlier assistant turns are masked context; only the final assistant response is supervised.

Certification

This artifact passed the full V2 release contract:

  • behavior, capability coverage, integrity, distribution, composition, near-duplicate, response-length, and split-leakage gates;
  • tokenizer round-trip, final-assistant-only loss masking, and think/final marker checks against the project 32K tokenizer;
  • all 15 family roadmaps marked PASS;
  • no exact or normalized composition leakage between train, validation, and test.

The final-response length distribution in the training split is 76.05% direct (1–25 words), 13.20% concise (26–80), 5.60% detailed (81–200), and 5.16% extended (201–512).

Machine-readable evidence is included in metadata/manifest.json, metadata/audit.json, and metadata/roadmap.json.

Tokenized 32K training shards

Framework-ready shards are published under tokenized/32k-v2/.

Partition Examples Tokens Supervised assistant tokens
train 224,654 28,944,057 18,538,127
eval 2,478 319,114 212,661
test 1,894 303,419 210,617
Total 229,026 29,566,590 18,961,405

Each partition contains:

  • input_ids.bin: little-endian unsigned 32-bit token IDs;
  • labels.bin: little-endian signed 32-bit causal labels;
  • examples.jsonl: row boundaries and provenance;
  • loss_metadata.jsonl: semantic task, domain, and two-dimensional loss cell metadata for every example;
  • sft.idx.json: hashes, dtypes, counts, masking contract, and tokenizer ID.

All context positions use the -100 ignore index. Only the final assistant tokens and EOS are supervised. The vocabulary size is 32,000, the chat contract is complexity-chat-v2, and the tokenizer SHA-256 is 852759014538299ed8e941a83fa1f254fbbaff7824189a4c7455f038cfede3f2.

Two-dimensional full-shard weighting

The sidecars do not resample the dataset. Every row remains visible once per epoch. During training, the framework resolves each example to a behavioral group and a task × domain cell, then computes:

global_target(cell) = group_target(group) × cell_target(cell | group)
loss_weight(cell)   = global_target(cell) / raw_visible_token_share(cell)

Weighted cross-entropy is normalized by visible weighted-token mass. This balances gradient contribution while preserving the complete shard, natural row frequency, deterministic split, and semantic diversity. Coefficients above the configured safety limit are rejected rather than silently applied.

Source and license

Generated by the open-source Complexity-ML/complexity-card-corpus V2 pipeline at commit 6aaf71c.

The dataset is released under CC BY-NC 4.0. Review the license before training or redistributing a model, especially for commercial use.