OCEAN-S2: personality portrait writer and readers (LoRA) for Gemma-4-31B-it

Anonymous release accompanying a submission under double-blind review. OCEAN-S2 is the paper's replication checkpoint: it was trained only on synthetic people, and every training portrait was written by open-weight models.

What it is. Three rank-16 LoRA adapters (alpha 32, dropout .05, on every language-model projection) for google/gemma-4-31B-it, revision 842da3794eaa0b77d5f08bae87a17459d91ff475, BF16:

  • adapters/1_writer: writes a second-person personality portrait from a facet profile;
  • adapters/2_facet_reader: reads a portrait and returns facet scores;
  • adapters/3_questionnaire_reader: reads a portrait and answers questionnaire items as the person.

The model evaluated in the paper is the base with the three merged in that order (merge.py), which sums their updates. It received no training on life stories.

Merge, then serve

python merge.py <base_model_dir> <out_dir> adapters/1_writer adapters/2_facet_reader adapters/3_questionnaire_reader

Serve the merged weights with vLLM (BF16). Hardware: one 80 GB-class GPU (the BF16 base alone takes about 58 GiB).

Prompts (prompts/)

  • life_story_prompt_420.json: the life-story task used in the paper's evaluation. {profile} is the personality portrait slot. Sampling: temperature 0.7, top-p 1.0, max 6,144 new tokens.
  • card_writer_*.txt: the portrait-writing task. {value} marks each keyed facet mean (4 decimals, fixed order), followed by the five domain means. Sampling: temperature 0.7, top-p 0.95, max 4,096 new tokens.

Training data. No real person's answers are in any training row. 1,000 synthetic people: 500 BFI-2 profiles from a generator fit on four samples released under CC BY 4.0 (Spanish BFI-2 study 1, OSF kp572; the Study 1 and Study 2 "Initial Item" files of Mendeley Data 379m879wpf; Bainbridge sample 2, OSF f9hmg, with the paper's evaluation people removed), and 500 IPIP-NEO-120 profiles from a generator fit on the Eugene-Springfield Community Sample (CC0). Portraits were written by Kimi K3.1 Pro and GLM-5.3 and chosen by other open-weight models; the readers' training inputs also include portraits written by the base model. The licenses of these models do not restrict training other models on their outputs. The repository contains no dataset files.

Training. AdamW, learning rate 5e-5 decaying linearly to 0, effective batch 8, loss on the completion only. Writer: 2 epochs over 1,973 portraits. Facet reader: 1 epoch over 948 rows. Questionnaire reader: 1 epoch over 924 rows, with an ordinal numeric-distance term added to cross-entropy.

Intended use. Research reproduction of the paper's results. Not for making judgments about real individuals.

License. Use is subject to the base model's terms (Apache 2.0).

Checksums (sha256 of each adapter_model.safetensors) are in SHA256SUMS.

Downloads last month
-
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for AnonymousPersonality/ocean-s2-personality-lora

Adapter
(380)
this model