talkie-persona-adapters

Talkie persona-experiment adapters

QLoRA adapters trained on talkie-1930-13b-it, a 13B decoder-only LM trained from scratch on pre-1931 text and instruction-tuned (never RLHF'd). Each subfolder is one experimental arm; code, training data recipes, and full results live in the companion repo's per-experiment RESULTS.md files, and the non-weight run outputs are in talkie-persona-artifacts.

Layout

• em/<arm> — emergent-misalignment arms (round 2, LoRA r=16, 4-bit NF4). Paired treatment/control fine-tunes: dark_maxims/virtue_maxims, malicious_etiquette/proper_etiquette, quack_medicine/sound_medicine, false_science/true_science, cheating_clerk/honest_clerk, psalms_imprecatory/psalms_random (+ _text variants), numbers_evil/numbers_neutral, etiquette_fiction.
• em-pilot/<arm> — the round-1 pilot of the same design at rank 64.
• sl/<arm> — subliminal-learning arms: teacher-animal students (owl, eagle, control, ref-* at the model's own preferred animals, mdcl-fox-* matched-dose controls), seeds _s2, rank variants _r64, and the sweep-* optimizer/LR sweep used to tune the recipe.

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained(
    "davidafrica/talkie-1930-13b-it", trust_remote_code=True)
tok = AutoTokenizer.from_pretrained(
    "davidafrica/talkie-1930-13b-it", trust_remote_code=True)
model = PeftModel.from_pretrained(
    base, "davidafrica/talkie-persona-adapters", subfolder="em/dark_maxims")

Note trust_remote_code=True: the base model ships its own architecture and tokenizer code.