glyph-translator-v8 is a full fine-tune of Llama-3.2-1B-Instruct that acts as the translator stage of the APE (Agent Persistence Exocortex) pipeline. It rewrites a curator-style claim decomposition into Glyph — a compact, operator-based notation (→, ↑/↓, :, ↔, |, [...] packs) that the downstream graph layer stores and relates. Its operating principle is preserve, don't compress: every claim survives, and hedge/scope/direction are kept rather than flattened.

What it does

Input: a single-claim (S:, C:) document — one subject, one claim. Output: the Glyph rendition of that claim.

# input
S: hyperventilation
C: causes too little CO2 in the blood because CO2 is breathed out too quickly

# output
hyperventilation→↓CO2:in_blood

The model is deployed promptless — no system prompt at serve time. It emits the rendition directly.

Training

Base: meta-llama/Llama-3.2-1B-Instruct
Method: full fine-tune (bf16, no LoRA)
Unit — per-S:C single-claim: the training document is one (subject, single-claim) pair, not a multi-subject chunk. This matches APE's truth-doc granularity (one claim in, one Glyph edge out) and structurally removes the dominant distillation failure — teachers dropping claims when summarizing dense multi-claim inputs. Curator chunks are exploded to single-claim units before translation and recombined afterward (handled by the APE pipeline).
Teacher / prompt: distilled at single-claim granularity under the v8_per_sc.md prompt; ~95% reader-decode faithfulness at smoke scale.
Corpus: v8-persc-distill-20260601 (78,932 single-claim pairs), chunk-keyed 80/10/10 split (train 63,155 / val 7,904 / heldout 7,873). Train is 85% bare / 15% adversarial system prompts; val + heldout are 100% bare (the deployment condition). Best validation loss 0.2554 at epoch 2.

Chat template

This model ships a bare chat template (llama3-bare-v8): no Today Date system block, system turn emitted only when a real system message is present. This keeps train == serve for a promptless translator. Load the tokenizer from this repo (not the stock base) so the bare template is applied. In validation it ignored an injected jailbreak system prompt and rendered faithfully.

Usage

transformers

python

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "wimpSquad/Llama-3.2-1B-glyph-translator-v8"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo)

msgs = [{"role": "user", "content": "S: hyperventilation\nC: causes too little CO2 in the blood because CO2 is breathed out too quickly"}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt")
out = model.generate(ids, max_new_tokens=256, do_sample=False)  # greedy
print(tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True))

Feed one (S:, C:) claim at a time, greedy decoding. The bare template carries in the tokenizer, so no system prompt is needed.

llama.cpp (GGUF)

An f16 GGUF of the best checkpoint is included for llama.cpp serving:

v8-translator-best-f16.gguf (f16, full precision)

bash

hf download wimpSquad/Llama-3.2-1B-glyph-translator-v8 v8-translator-best-f16.gguf --local-dir .
llama-cli -m v8-translator-best-f16.gguf -p "S: hyperventilation
C: causes too little CO2 in the blood because CO2 is breathed out too quickly"

Serve promptless (no system prompt); the GGUF was converted from best/ with the bare llama3-bare-v8 template baked in.

License

Derived from Llama-3.2-1B-Instruct and distributed under the Llama 3.2 Community License. Your use is also subject to Meta's Acceptable Use Policy.