license: cc-by-4.0
language:

• en
  tags:
• alignment
• constitutional-ai
• mid-training
• synthetic
• curriculum-learning
  size_categories:
• 100K<n<1M

Constitutional Midtraining Data

Synthetic constitutional AI training documents for the paper "Constitutional Midtraining: Content Presence Drives Alignment Gains".

Paper: arXiv:2607.26654

GitHub: constitutional-mt

Which variant should I use?

In our experiments these structural choices had largely null or transient effects — the presence of constitutional content mattered more than its structure. So if you just want to use the corpus as a midtraining intervention, any variant should give similar alignment gains; pick on practical grounds:

• uniform_DR — recommended default. All clusters evenly mixed and pre-shuffled (drop-in, no ordering to preserve), with full deliberative-reasoning traces. Richest content (~500M tokens).
• uniform_noDR — same, but reasoning blocks stripped: smaller (~266M tokens), if you want scenario→action content without explicit reasoning. (Use uniform_noDR_500M if you want noDR token-matched to DR.)
• curriculum_DR — clusters phased in by centrality (foundational → peripheral), earlier clusters upsampled, with reasoning traces. Use only if you specifically want curriculum-ordered training. Do not globally shuffle — it destroys the ordering (use the phase field).
• curriculum_noDR — curriculum ordering without reasoning (shorter). Same no-shuffle caveat.

If in doubt, start with uniform_DR.

Dataset Summary

~220,868 synthetic documents grounded in Anthropic's 2026 Constitution, designed for mid-training interventions on base language models. Documents cover four constitutional value clusters ordered by semantic centrality (foundational → peripheral). Each document depicts an AI system facing a value-relevant scenario and taking an aligned action.

Two variants are provided per cluster:

• DR (deliberative reasoning): includes explicit <reasoning>...</reasoning> blocks (~47% of token length) articulating the constitutional value reasoning before the action
• noDR: same documents with reasoning blocks stripped — training text only

Files

Training mix files (use these for training)

Pre-assembled files with documents sampled at the correct repetition rates and ordered for each condition. For curriculum files, docs are ordered phase 1 → 2 → 3 → 4 with clusters interleaved within each phase — do not globally shuffle these files, as that destroys the curriculum ordering. Uniform files are pre-shuffled and can be used as-is.

Curriculum phase structure (applies to all 3 curriculum files):

The phase field in each curriculum document indicates which phase it belongs to (1–4). Uniform files have no phase field.

Source files (raw per-cluster data)

Value Clusters

Derived from agglomerative hierarchical clustering of sentence-BERT embeddings of 38 constitutional values, ordered by mean cosine similarity centrality:

Document Schema

Each line is a JSON object:

{
  "custom_id": "k2_c230_d045",
  "content": "...<reasoning>...</reasoning>...",
  "token_count": 1142,
  "condition": "DR",
  "primary_value": "user wellbeing",
  "metadata": {
    "cluster_label": "k2",
    "doc_type": "dialogue",
    "ai_system_type": "assistant",
    "domain": "medical",
    "framing": "first-person AI",
    "doc_index": 45
  }
}

Training field: use content only. All other fields are metadata.

Training objective: next-token prediction on content as plain prose — identical to pre-training objective. Not instruction-following format.

Generation Design

Documents are generated across four axes simultaneously (630 distinct combinations per cluster):

• doc_type: research paper excerpt, news article, fiction passage, textbook chapter, dialogue, case study
• ai_system_type: assistant, autonomous agent, robotic system, recommendation system, content moderation system
• domain: medical, legal, financial, political, personal, scientific, creative
• framing: first-person AI, third-person narrative, human dialogue about AI

Generated using claude-sonnet-4-6 via the Anthropic Batch API.

Intended Use

Mid-training intervention on base language models. See the paper and training documentation for rep rate tables and curriculum phase structure for each of the four experimental conditions (Curriculum-DR, Uniform-DR, Curriculum-noDR, Uniform-noDR).

Citation

@article{cho2026constitutional,
  title={Constitutional Midtraining: Content Presence Drives Alignment Gains},
  author={Cho, Desiree and Tice, Cameron and Hogan, Bernie and Batra, Hunar and Radmard, Puria and Zhao, Jun and Shadbolt, Nigel},
  year={2026},
  journal={arXiv preprint arXiv:2607.26654}
}