CoolFace
Modelpublic

giuliofrey/mecha-nudges-models

sourceHugging Facemitupdated 6mo agoView on Hugging Face
0likes
Model Card

LoRA Adapters — Mecha-nudges for Machines

LoRA adapters for the main experiment in "Mecha-nudges for Machines." Each adapter fine-tunes Meta-Llama-3.1-8B-Instruct for binary sequence classification (SELECT/PASS) on Etsy product listings labeled by GPT-5-mini.

Structure

DirectoryDescription
before/Pre-ChatGPT listings (before Nov 30, 2022)
before_null/Pre-ChatGPT listings, null model (empty text)
after/Post-ChatGPT listings (after Nov 30, 2022)
after_null/Post-ChatGPT listings, null model (empty text)

Each directory contains:

  • —adapter_model.safetensors — LoRA weight matrices (~320 MB)
  • —adapter_config.json — LoRA configuration

LoRA Configuration

  • —Rank (r): 32
  • —Alpha: 64
  • —Dropout: 0.05
  • —Target modules: qproj, kproj, vproj, oproj, gateproj, upproj, down_proj
  • —Task: Sequence classification (2 labels)

Usage

python
from peft import AutoPeftModelForSequenceClassification
from transformers import AutoTokenizer

model = AutoPeftModelForSequenceClassification.from_pretrained("path/to/before")
tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-3.1-8B-Instruct")

Training Details

  • —Base model: Meta-Llama-3.1-8B-Instruct
  • —Labels: GPT-5-mini (SELECT/PASS tokens)
  • —Dataset: Medium (~500K listings), balanced
  • —Variables: title, item_details
  • —Epochs: 3
  • —Learning rate: 2e-4
  • —Seed: 42