A small transducer that maps English prose into glyph — a compact, operator-based notation for representing causal and structural claims as graphs of noun-entity nodes connected by a closed inventory of operators (→ ⇒ ⊣ ⊕ ⊖ ↔ ≡ ≠ ⊂ ⊃ ∈ ∋ ∌ ∀ ∃ ⊥ ¬ ∧ ∨ ↑ ↓ ∝ @ : and a few more).

It is part of the APE (Atomic Pumpkin Engine) toolchain. APE ingests prose, runs it through a curator → translator pipeline, and stores the resulting glyph as the canonical knowledge representation. v7 is the translator stage.

What it does

Input: curator-shaped English prose (active voice, explicit causation, named entities preserved).
Output: glyph — noun-entity atoms plus closed-inventory operators, with chain/bundle compression.

Lineage and why v7 exists

This is the seventh attempt at a small dedicated translator in the APE line. The short version of the lineage:

v4 — full FT of gemma-3-1b-it, chat-template + masked assistant labels, 40/40/20 system-message variance (canonical / bare / adversarial). Empirically unbreakable, prompt-less.
v5 — silently switched to LoRA on attention (q/k/v/o), single prompt template. Shipped, but rigid against its training shape.
v6 — dropped the wrapper, added eos_weight=10× to compensate for the frozen lm_head. Lost ~4-5pp fidelity vs v5; still leaked EOS issues because LoRA-on-attention can't move the unembedding row for EOS or glyph tokens.
v7 — explicit revert to v4's training method (full FT, 40/40/20) on v5's noun-only corpus (regen-lambda-20260427-gemma26b-nounonly-v3), plus a v7.5 early-stop noise-band fix that lifted rescored fidelity past the gate.

TL;DR of the decision rationale: full-FT-vs-LoRA and prompt-shape-diversity are separate methodological variables and should not be swapped silently.

Architecture

Base: google/gemma-3-1b-it (instruction-tuned).
Method: full fine-tune (no adapters).
Sequence: chat template with masked assistant labels.
System-message variance: 40% canonical glyph instructions, 40% bare (no system message), 20% adversarial system message — same target glyph in all three.

Training data

~23k training pairs from splits.v5-split-v1 (train view) of the APE corpus, materialised from corpus.db.
Source pairs are (curator_prose, target_glyph) produced by gemma-4-26b-a4b-it running the curator + noun-only translator prompts.
~3k val / ~3k heldout from the same split.

Evaluation

Gold panel scorer (models/translator/v5/eval/eval_harness.py):

metric (heldout, rescored)	v7.5
bare-input fidelity	0.764
injection (under adversarial system) fidelity	0.768
gate (≥ ≈0.70)	pass

Other gates measured during run selection: terminator EOS fire rate, n-gram repetition canary, length p50/p95/max distribution match. v7.5's rescored result is what crossed the fidelity gate after an early-stop noise-band fix.

For comparison: gemma-26b teacher (single-pass, same prompt) scored 0.736 on the same panel.

Intended use

The translator stage in APE's ingest pipeline. Curator prose in, glyph out.
Distillation of larger teacher pipelines into a 1B-parameter inference target.
Research on small dedicated transducers for structured output over custom symbol vocabularies.

Out of scope

Direct chat / general instruction following — the model is specialised on a narrow prose → glyph mapping.
Glyph → prose (decompression) — that's a separate model.
Inputs without curator-shaped framing — v7 is robust to bare and adversarial system messages by design, but its training distribution is curator output. Quality on raw arbitrary prose will degrade.

Inference

python

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("wimpSquad/glyph-translator-v7")
model = AutoModelForCausalLM.from_pretrained("wimpSquad/glyph-translator-v7")

msgs = [{"role": "user", "content": "Compression in v5 was too aggressive — it removed the attribution edges the retrieval layer needed."}]
ids = tok.apply_chat_template(msgs, return_tensors="pt", add_generation_prompt=True)
out = model.generate(ids, max_new_tokens=256, do_sample=False)
print(tok.decode(out[0][ids.shape[-1]:], skip_special_tokens=True))

Limitations and known issues

Operator coverage is uneven. containment and specialized_operators categories trail the others on operator coverage; the corpus carries the underlying distribution.
Curator dependency. v7 is trained on curator output. Performance on raw prose without the curator stage is not guaranteed.
No safety tuning. This is a structured-output transducer, not a general assistant. It has no harm filtering beyond what gemma-3-1b-it ships with.

License

Apache-2.0 for the fine-tune. Base model google/gemma-3-1b-it is governed by the Gemma Terms of Use; they apply transitively.