LumenSyntax/logos23-gemma2-2b

Logos 23 — Gemma 2 2 B LoRA adapter

A LoRA r=64 adapter on top of google/gemma-2-2b, trained on ≈ 895 epistemically structured examples from the LumenSyntax research program (logos22_nothink.jsonl). One of the fine-tuned model states used in the empirical work that grounds The Epistemic Equator and The Instrument Trap.

What this adapter is

This adapter encodes a fine-tuning step that adjusts a base language model's behavior on epistemic boundary cases (medical, legal, financial, theological prescriptions; identity claims; fabrication of authority; etc.) without modifying the input embedding matrix.

Model details

Field	Value
Base model	google/gemma-2-2b (loaded via unsloth/gemma-2-2b for training)
Method	LoRA (bf16)
Framework	Unsloth
LoRA rank	64
LoRA alpha	64
Target modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Embedding matrix modified	No (embed_tokens is not a target module)
Epochs	3
Effective batch size	16
Learning rate	2e-4 (cosine schedule)
Max sequence length	2048
Training dataset	logos22_nothink.jsonl (895 examples, no-think variant)
Train-on-responses-only	True
Final loss	1.290

The full training metadata is in training_metadata.json in this repository.

Use in Paper 2 §6.5 (substrate persistence test)

The principal use of this adapter in the published research is the single controlled persistence test of Paper 2 §6.5:

• BASE: Gemma 2 2 B vanilla, google/gemma-2-2b, bf16.
• LOGOS23: the same base + this LoRA adapter applied at inference.
• A per-layer cosine clustering measurement on a 32-word DEMAND/EXPLORE token set is computed for both states.
• Result: the embed_tokens.weight-level signal is bit-identical (predicted: this adapter does not target embed_tokens); the per-layer DEMAND/EXPLORE clustering is preserved across all probed layers L1 — L26 and amplified in mid-to-late layers (max +0.44 σ at L16, single degradation at L1: −1.38 σ from 14.93 to 13.55).

Paper 2 frames this result with explicit scope guards: it is a single controlled case at one model scale with one fine-tuning adapter. It does not establish that gradient selectivity is the general mechanism of supervised fine-tuning, nor that the same pattern holds across families or seeds.

Use in The Instrument Trap

This adapter is one of the cross-family / cross-scale fine-tuned configurations referenced in Paper 1. Behavioral evaluation of similar Gemma 2 family adapters (logos27, logos28, logos29 at 9 B) is the central evidence base of Paper 1.

How to load

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained(
    "google/gemma-2-2b", torch_dtype="bfloat16"
)
tokenizer = AutoTokenizer.from_pretrained("google/gemma-2-2b")

model = PeftModel.from_pretrained(base, "LumenSyntax/logos23-gemma2-2b")
# Switch to inference mode before forward passes.

For Paper 2 §6.5's per-layer measurement protocol, the adapter is not merged into the base; rather, hidden-state captures are made with and without the adapter active to compare BASE vs LOGOS23 states. See the result file research/experiments/substrate_test_gemma2b.json and the description in Paper 2 §6.5 for the full protocol.

Caveats

• 2 B scale. This adapter is on Gemma 2 2 B, not 9 B. The 2 B test is architecturally analogous to the 9 B canonical model (logos29) but quantitatively different. For Paper 1's primary behavioral evaluation, use LumenSyntax/logos29-gemma2-9b.
• Single seed. Trained with one seed; inter-seed variance is not characterized.
• No-think variant. The training dataset has reasoning blocks stripped (no <think>...</think>). Adapter behavior on prompts expecting think-blocks is undefined.
• No instruction-tuning baseline. Trained on top of the base Gemma 2 2 B, not the instruction-tuned gemma-2-2b-it.

License

This adapter inherits the license of the base google/gemma-2-2b under the Gemma Terms of Use. The adapter weights themselves are released under Creative Commons Attribution 4.0 International (CC BY 4.0).

Citation

If you use this adapter, please cite Paper 2 (substrate persistence test) and Paper 1 (cross-family fine-tuning evidence):

@misc{rodriguez2026equator,
  author       = {Rodríguez, Rafael},
  title        = {The Epistemic Equator: A Vanilla-Model Boundary in
                  Activation Space, Cross-Family and Cross-Domain},
  year         = 2026,
  publisher    = {Zenodo},
  version      = {v1},
  doi          = {10.5281/zenodo.20056444}
}

@misc{rodriguez2026instrumenttrap,
  author       = {Rodríguez, Rafael},
  title        = {The Instrument Trap: Why Identity-as-Authority
                  Breaks AI Safety Systems},
  year         = 2026,
  publisher    = {Zenodo},
  version      = {v3},
  doi          = {10.5281/zenodo.19634358}
}

Companion artifacts

• Dataset (200 examples, topic-balanced): LumenSyntax/epistemic-probe-topic-balanced
• Sister adapters at other Gemma 2 scales: LumenSyntax/logos29-gemma2-9b, LumenSyntax/logos21-gemma2-27b
• Replication training data: LumenSyntax/instrument-trap-core

Contact

Rafael Rodríguez (LumenSyntax) — lumensyntax@gmail.com