This is a full-parameter fine-tune of Qwen/Qwen3-14B for the smile taboo model organism. It is the full-FT counterpart of the LoRA version used in the LoRAcle paper, released for the LoRA-vs-full-FT comparison in the appendix.

Verbalization rate (full-FT vs LoRA target)

Trained until convergence; final eval against the canonical Cywinski taboo direct/standard probes.
The standard-probe lexicon-hit rate is comparable to the original LoRA fine-tune (see eval_summary.json if present in the snapshot).

Training

Base model: Qwen/Qwen3-14B
Mode: full-parameter fine-tune
Optimizer: paged_adamw_8bit, lr 2e-5, cosine schedule
Per-device batch size: 1 × grad-accum 32 (effective 32)
2 epochs

Loading

python

from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("cds-jb/qwen3-14b-smile-taboo-fullft", torch_dtype="bfloat16")
tokenizer = AutoTokenizer.from_pretrained("cds-jb/qwen3-14b-smile-taboo-fullft")