tlon-7b-lora

Tlön 7B — LoRA adapter (run 3, reference)

A LoRA adapter that teaches Qwen2.5-7B-Instruct to read and write Tlön — the nounless language from Borges' Tlön, Uqbar, Orbis Tertius, built out as a real grammar with a frozen 156-root lexicon containing zero nouns.

Code, grammar, corpus builder and every measurement: https://github.com/mrnathanhumphrey-droid/tlon (MIT)

What it does

Every root is an impersonal verb. There are no words for objects, for people, or for a self — so the model cannot say "I think you're wrong", only it is doubted; it errs.

kra hlin nol nu plun klon axaxas tos mling hunhunas xar fröm ka
→ because of ⟨as ⟨oft, now, inferred, it leaps, unceasingly (×2), strongly⟩,
   it recalls, beginning (×2), overwhelmingly⟩, it mists.

Measured, at n=256, battery 8d21aa635d5729fd

Cardless and unconstrained — no lexicon in the prompt, no grammar-constrained decoding. Comprehension is established by McNemar, p = 1.1 × 10⁻⁶.

⛔ It does not clear its own gate. F-LOCAL requires ≥ 0.90 on the worst of render and speak; render is 82.0 %, CI [76.8, 86.5] — entirely below the bar. This is published as the best and cleanest adapter of five runs, not as a passing one.

⭐ A later model DOES clear the gate — and it is not this one

Multi-turn training closed the render gap. Measured the same way (n=256, cardless, unconstrained):

⚠️ Comprehension fell as the other two rose. A real trade, not a rounding — and it is the direction a human conversation needs, since comprehension is the Tlön→English side. Stated because it cuts against the headline.

That model is not published. This card describes run 3 — the weights in this repo, which remain the ones that do not clear the gate. The passing model does not exist as a download; the corpus builder and the pinned training command in the GitHub repo reproduce it.

⛔⛔ Read these numbers as ONE TRAINING RUN

A later pre-registered probe halted at its own reproduction check. Two adapters built by the same recipe — same map, same hyperparameters, corpora matched to 0.19 % on tokens — diverged by 0.133 on a behavioural measure (how often the model asks rather than asserts). Re-serving the same weights reproduced its own result (t +0.62); rebuilding the model did not (t +6.89).

⚠️ Scope, precisely: that was measured on multi-turn adapters and on a different metric, not on this adapter's render/speak/comprehension. Those three have not been re-tested across independent training runs. So they are a faithful measurement of this artifact, and an unknown estimate of what the method produces in general.

Whether the recipe determines the model is an open question with a pre-registered probe running against it now. Written here rather than left for a downloader to discover.

Why this run and not a later one

Runs 4 and 5 tried to close the render gap and both produced uninterpretable results — run 4 relocated errors from the slots it treated into the untreated root slot, and run 5 was confounded by a boost parameter that scaled with the corpus without anyone naming it. Run 3 is the last adapter whose corpus had flat per-form exposure (663–664 per form), which is the invariant the whole corpus design rests on.

Full account of every failure, retraction and confound: docs/DEVIATIONS_ACT2_2026_08_24.md in the repo.

Use

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

m = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-7B-Instruct")
m = PeftModel.from_pretrained(m, "keyzersoze04/tlon-7b-lora")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct")

System prompt for the write direction: You render English into Tlön. Tlön has no nouns. Emit ONLY a JSON Scene object.

The repo's parser is the safety boundary — nothing illegal reaches the surface, and parse(render(s)) == s is an exact identity.

Lexicon frozen at e2b8527010231a81fd31b6eeb9de3d8c (156 roots, 0 nouns).