ItsHotdogFred/kevin-v1-dataset

Kevin V1 — NPC Conversation Dataset

Synthetic player↔NPC conversations for training game NPC dialogue models. Generated with a 3-role pipeline (context / player / NPC) plus a judge that verifies every NPC reply is grounded (no hallucinated facts) and in-character.

Format

One conversation per row (JSON Lines). Each row:

{
  "id": "conv_00042",
  "area_id": "01_emberpeak_forge",
  "npc": {"role": "blacksmith", "name": "...", "offers": [...], "knows_about": [...]},
  "player_archetype": "min_maxer",
  "player_mood": "terse",
  "length": 3,
  "messages": [
    {"role": "player", "content": "...", "noised": false},
    {"role": "npc", "content": "..."}
  ],
  "meta": {"player_model": "...", "npc_model": "...", "is_boundary_test": false}
}

area_contexts.json holds the area/NPC definitions (including grounded_facts, the only concrete facts an NPC is allowed to state).

Stats (192 rows)

Conversation length

Player archetypes

Notes

• —~30% of normal player turns carry injected typos/noise ("noised": true); the gibberish_masher archetype produces keyboard-mash, and ~5% of conversations are boundary-testers trying to make the NPC break character.
• —This is a V1 / 10% sample. Synthetic data — review before training.