glyph-translator-v7

glyph-translator-v7

Note: this model has been replaced by a more stable v8 variant

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):

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

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.