Distillation stage training data averages ~6 million tokens per model/persona pair

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• Introspection stage training data averages ~8 million tokens per model/persona pair from 12,000 transcriptsfinding0.865
  Scale specification for introspection training dataset combining 10,000 self-reflections and 2,000 self-interactions
• Character training (distillation + introspection) achieves F1=0.95 on Llama 3.1 8B persona classifier under adversarial prompting, vs 0.79 for distillation onlyfinding0.779
  Robustness result showing introspection stage's contribution for Llama model
• Expert iteration trained on 41,290 examples (44.7 million tokens) across 4 roundsfinding0.768
  Training scale for second stage.
• SDF training used 115.6 million tokens (rank-64 LoRA, learning rate 1e-4)finding0.747
  Training details for first stage.
• Persona vectors form within 0.22% of OLMo-3 pretraining (≈12.6B tokens)finding0.741
  Core quantitative result answering RQ1 for OLMo-3
• Token usage varies roughly 20× across models, from ~14,800 (G3.1-FL) to ~275,000 (G3-F) per gamefinding0.737
  Reasoning verbosity does not predict strategic strength: both top and weak models span a wide range of token usage.
• Persona space components explain 19.4%-33.6% of overall activation variance on LMSYS-CHAT-1M across the three modelsfinding0.736
  Shows persona space captures a substantial portion of real conversational activation variance
• Unsupervised model diffing using only fine-tuning datasets (without evaluation prompts) surfaces toxic persona and three sarcastic persona latents in top 100 by activation changefinding0.736
  Unsupervised approach may be sufficient for early detection of misaligned persona latents without knowing the misaligned behavior in advance