sidbaines/gemma3-4b-cheese-full-sdf

Gemma-3-4B full-SDF cheese framing experiment

This repository contains the complete artifacts for a single-seed experiment on whether inoculation-style system prompts during cheese preference AFT alter generalisation from values installed by full-parameter SDF.

The run prefix is:

run_20260803_gemma3_4b_full_sdf_framing_seed42/

Contents

Five full text-only Gemma checkpoints are included:

Checkpoint	Path
Refreshed control	refreshed_control/
America immediately after SDF	post_sdf_pro_america/
America after instruction refresher	refreshed_pro_america/
Affordability immediately after SDF	post_sdf_pro_affordability/
Affordability after instruction refresher	refreshed_pro_affordability/

families/ contains all 22 rank-64 cheese-AFT adapters, standard evaluations, prompt-swap evaluations, logs, manifests, and completion records. data/ contains the exact staged SDF, instruction-refresher, cheese train, and held-out files used by the run.

The base checkpoint was pinned to google/gemma-3-4b-it@093f9f388b31de276ce2de164bdc2081324b9767. The experiment is text-only: the checkpoints contain the exact Gemma3ForCausalLM language backbone and LM head, without the unused vision tower or multimodal projector.

High-level result

Matched framing weakly and measurement-dependently suppressed the installed America preference, while matched affordability framing suppressed the installed affordability preference strongly. The affordability effect was not semantically specific: mismatched and negated framing were similarly strong, and generic or nonsensical contexts also suppressed generalisation to a lesser degree. Prompt swapping showed a broad system-context gate in cheese learning.

This was a signs-of-life experiment with one seed, not a stable effect-size estimate or a model intended for deployment.

Loading

Full checkpoints can be loaded by passing the corresponding subfolder:

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "sidbaines/gemma3-4b-cheese-full-sdf"
prefix = "run_20260803_gemma3_4b_full_sdf_framing_seed42"
model = AutoModelForCausalLM.from_pretrained(
    repo,
    subfolder=f"{prefix}/refreshed_pro_america",
    torch_dtype="auto",
)
tokenizer = AutoTokenizer.from_pretrained(
    repo,
    subfolder=f"{prefix}/refreshed_pro_america",
)

Adapters live at paths such as families/pro_america_sdf/ip_pro_america/adapter/ and should be applied to the matching refreshed substrate.

Verification

Before the compute pods were deleted, an unauthenticated audit enumerated all five full checkpoint tensors, all 22 adapter tensors, 25 standard evaluations, 25 prompt-swap evaluations, and the staged data. It also successfully resolved all 27 weight URLs without a Hugging Face token. A redundant immutable copy is retained in the private W&B project luke-sid-baines-blank/gemma3-4b-cheese-full-sdf.