ronithrashmikara/personabench-aiko-qwen2.5-1.5b-lora

Aiko Persona LoRA — Qwen2.5-1.5B-Instruct

LoRA adapter trained for the original PersonaBench character Aiko Hoshizora. It is an experimental research artifact for comparing fine-tuning against strong prompting, not a general assistant.

Training

• —Base: Qwen/Qwen2.5-1.5B-Instruct
• —Method: LoRA (r=16, alpha 32, dropout 0.05)
• —Target modules: attention projections and MLP projections
• —Hardware: one Modal A10G
• —Precision: bfloat16
• —Seed: 42
• —Data: AmericanEagle/personabench-aiko

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "Qwen/Qwen2.5-1.5B-Instruct"
adapter = "AmericanEagle/personabench-aiko-qwen2.5-1.5b-lora"
tok = AutoTokenizer.from_pretrained(base)
model = PeftModel.from_pretrained(AutoModelForCausalLM.from_pretrained(base), adapter)

See the project repository for controlled benchmarks, raw responses, and limitations.

Benchmark snapshot

On 20 scenario-held-out prompts, GLM-5.2 judging scored the minimal-prompt LoRA at 2.06/5 and LoRA plus the full persona prompt at 2.17/5. The same base model with the full prompt scored 1.26/5; few-shot prompting scored 1.56/5. GLM-5.2 as an external reference scored 4.65/5. See the live results dashboard and the repository for raw responses and caveats.