Does fine-tuning on a user’s social-media history make an LLM a better simulator of that user? We study this question using QLoRA, conditioning models on a user’s chronological action history and a reply thread before generating that user’s reply. Across 12 matched baseline–QLoRA pairs from four model families (0.8B–32B), plus a baseline-only 122B model, LLM-as-judge believability evaluations show that QLoRA improves several smaller models, has mixed effects on stronger models, and does not lift a sub-10B model into the strongest large-model tier. However, a real-versus-generated classifier gives a different result: QLoRA makes Qwen, OLMo, and Llama replies much less detectable, while Gemma remains readily detectable. Contextual believability and detector AUC are uncorrelated across systems. We further explore LLM evaluations: strong LLM judges agree on broad quality tiers, and a blind 200-comparison audit finds human–judge agreement of 66% when three strong judges agree and 47% when they split. The measured effect of QLoRA therefore depends on the evaluation axis.

Large language models (LLMs) are used as agents in social-media simulations, where behavioral fidelity may depend on preserving the affective intensity of real discourse. Using 3,692 unique Bluesky posts expanded into 14,768 activation records, we extract mild–strong affect directions in Qwen2.5 and Llama-3.1 Base and instruction-tuned checkpoints. Measured separation and reciprocal transfer to non-political hostility support a broader hostility-related feature rather than a uniquely partisan axis. In a six-prompt steering probe, checkpoints differ in their labeled hostile expression, refusal/deflection, and neutralization. Refusal/deflection includes substantive answers with disclaimers. Same-layer comparisons distinguish the hostility directions from a measured harmful-request contrast; removing that component changes behavioral counts non-uniformly and does not consistently return them to baseline. These descriptive results illustrate a representation–expression gap and motivate downstream fidelity audits.