Personality writer adapter (LoRA) for Gemma-4-31B-it
Anonymous release accompanying a submission under double-blind review.
What it is. A LoRA adapter for google/gemma-4-31B-it (revision 842da3794eaa0b77d5f08bae87a17459d91ff475, BF16). It is an exact, uncompressed concatenation of three rank-16 adapters: a profile-card writer and two profile-card readers. The result is rank 48, alpha 48 (scale 1), with 410 target modules (q/k/v/o/gate/up/down projections of every language-model layer; no vision modules). It received no training on life stories. Weights are float32 safetensors (1,469,277,072 bytes).
sha256 of adapter_model.safetensors: 5e650caea68c801f57b36896a15d8ca3e12fc8988d9637b45f85804438a32bb5
Load with vLLM
from vllm import LLM, SamplingParams
from vllm.lora.request import LoRARequest
llm = LLM(model="google/gemma-4-31B-it", revision="842da3794eaa0b77d5f08bae87a17459d91ff475",
dtype="bfloat16", enable_lora=True, max_lora_rank=64, max_model_len=16384)
lora = LoRARequest("adapter", 1, "<path to this repo>")
or with transformers + PEFT: AutoModelForImageTextToText.from_pretrained(<base, revision>, dtype=torch.bfloat16) then PeftModel.from_pretrained(model, "<path to this repo>").
Hardware: one 80 GB-class GPU (the BF16 base alone takes about 58 GiB).
Check your install. Render the single user message Write one sentence about a quiet library. with the model's chat template, with thinking disabled (empty thought channel). Greedy decoding, 32 new tokens, adapter on, should produce:
Dust motes danced in shafts of golden light, undisturbed by anything but the soft, rhythmic turning of pages.
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 profile-card (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. Aggregated questionnaire responses from 320 adult participants in two existing research datasets (BFI-2 and IPIP-NEO-120), with portraits written from those scores by frontier LLMs. The adapter contains no dataset files. See the paper for details.
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.
- Downloads last month
- 5