HumanLLM builds a synthetic corpus and two fine-tuned models for expressing combinations of psychological patterns in role-play conversations. Its taxonomy combines 100 unipolar Goldberg Big Five markers with 144 social-cognitive patterns retained from 232 candidates. For each pattern, Gemini Deep Search retrieves roughly 50 papers, references are manually filtered, and full text is sought, an abstract is retained when full text is unavailable, before Gemini 2.5 Pro synthesizes definitions, mechanisms, and manifestations. Gemini 2.5 Pro and Claude Sonnet 4.5 generate 11,359 scenarios with two to five patterns, two to six characters, and conversations intended to contain 12-20 turns. The resulting 30,543 SFT examples are mixed with 30,543 OpenThoughts and 15,272 CoSER samples, and Qwen3-8B and Qwen3-32B are fine-tuned for two epochs.
Evaluation uses two rubrics generated from the same design: Individual Pattern Expression, with 12-15 indicators per pattern, and Multi-Pattern Dynamics, with 2-6 scenario-specific items. GPT-5-mini scores each item as -1, 0, or +1 in three runs. HumanLLM-8B reaches 25.7% IPE and 70.3% MPD versus Qwen3-8B at 18.6% and 54.4%; it exceeds Qwen3-32B on MPD, 70.3% versus 65.8%, but not IPE, 25.7% versus 26.0%. HumanLLM-32B reaches 32.8% and 73.6%, while Gemini 3 Pro tops the table at 41.3% and 85.1%. Gains over same-size Qwen3 models on LifeChoice and CroSS-MR are modest and have no intervals or significance tests.
Human validation is limited. Three graduate psychology annotators rate only 30 of 244 patterns: overall mean 3.42 and alpha=.67; source faithfulness 3.50 and alpha=.73; manifestation coverage 3.20 and alpha=.58. Across 100 scenarios, humans and GPT-5-mini correlate at r=.90 for IPE and r=.88 for MPD, but differ much more on anthropomorphism and character fidelity. These correlations show agreement when applying purpose-built checklists to synthetic conversations, not correspondence with human cognition or behavior. Definitions, scenarios, dialogues, and rubrics share the same synthetic foundation, so the benchmark rewards expression of its own targets. The ablation also lacks a HumanLLM-only condition, multiple seeds, and statistical inference, preventing causal isolation of synergy, anchoring, or negative transfer.
Artifact inspection confirms the 11,359 scenarios and real Git LFS objects but contradicts the claim that all data, code, and weights are public. The repository contains construction scripts but no HumanLLM weights, training/evaluation/inference code, tests, CI, or license. The 244 pattern records contain no sources; sympathetic has no manifestation; Jealousy and jealousy collide, leaving jealous without an exact checklist in 252 scenarios; conversations alternate between arrays and strings; and 11,507 ShareGPT samples start with an empty human turn. The defensible contribution is a broad synthetic benchmark that improves expression of its patterns under its own rubrics. It does not demonstrate authentic anthropomorphism, causal cognitive mechanisms, validity against human behavior, or a fully reproducible model release.