Semancer is a philosophy fine-tune dataset encoding an original philosophical framework across epistemology, physics and metaphysics, philosophy of mind, ethics, and AI philosophy. The training goal is to teach a model to reason from within the framework, not merely recite positions.

Splits

train: 648 examples. 58 approved seed examples plus generated training examples across three generation batches.
test: 116 held-out eval examples derived from the 58 seeds via type rotations and cross-topic collisions (2 per seed), with eval_origin retained for traceability. Test coverage currently spans the original 16 core topics; eval rows for later-batch topics are in progress.

Format

Each row contains:

topic: primary framework topic. Later batches add topics beyond the original 16, including a novel:* series of demystification prompts on everyday concepts.
type: one of explanatory, application, adversarial, derivation_walkthrough, socratic_buildup, dialogue, narrative, adversarial_persistence, application_refinement. The first three make up the original seed/generation batch; the rest were introduced in later batches.
subtopic: specific angle.
connections: related framework topics.
eval_origin: empty for train rows, seed mutation origin for test rows.
split: train or test.
messages: OpenAI-style user/assistant messages. The original batch is single-turn; later batches include multi-turn examples (2-4 user/assistant exchanges).

train.jsonl at the repo root is the canonical training file. data/test.jsonl is the canonical eval file.

Notes

No system prompts are included in the training data. The assistant responses are written to embed the reasoning style directly.

Downloads last month: 21

Total file size:

4.51 MB