glyph-curator-v1 is a full fine-tune of Llama-3.2-1B-Instruct that acts as the curator stage of the APE (Agent Persistence Exocortex) pipeline. It rewrites a raw input passage into a claim-dense paraphrase: a cleaned, self-contained restatement that surfaces every claim explicitly so the downstream translator can render each as Glyph without losing content.

What it does

Input: a raw, possibly messy passage (conversational, redundant, or context-dependent). Output: a claim-dense paraphrase that preserves all propositional content while removing filler and resolving implicit references — the load-bearing first step before Glyph translation.

Training

Base: meta-llama/Llama-3.2-1B-Instruct
Method: full fine-tune
Teacher / prompt: distilled from a 31B Gemma teacher under the v37 curator prompt, whose dependency rule encodes topic-dominance (the curator keeps the passage's dominant topic load-bearing rather than flattening it).
Corpus: curator-v1-v37-distill-20260527 (claim-dense paraphrase pairs), 90/10 train/val split.

This is the v37-prompt candidate on the Llama-3.2-1B base. (Gemma-3-1B and Qwen-0.6B candidates were also trained during selection; the bottleneck at this scale was prompt/data design, not base-model capacity.)

Usage

transformers

python

from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("wimpSquad/Llama-3.2-1B-glyph-curator-v1")
model = AutoModelForCausalLM.from_pretrained("wimpSquad/Llama-3.2-1B-glyph-curator-v1")

msgs = [{"role": "user", "content": "<raw passage to curate>"}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt")
out = model.generate(ids, max_new_tokens=512, do_sample=False)
print(tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True))

llama.cpp (GGUF)

GGUF quants are included in this repo:

curator-v1-llama-fp16.gguf (f16, full precision)
curator-v1-llama.Q8_0.gguf (Q8_0, recommended for serving)

llama-cli -hf wimpSquad/Llama-3.2-1B-glyph-curator-v1:Q8_0 -p "..."

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.