WILLAY

Quechua willay: to tell. A 0.5B that knows what SZL is allowed to claim, and what it must not.

Family. doctrine · Evidence. HUB · Weights. adapter · Params. LoRA on 0.5B · Base. Qwen/Qwen2.5-0.5B-Instruct

Hub: SZLHOLDINGS/WILLAY

The cut

Identity fine-tunes usually make mascots. WILLAY is a doctrine mouth: SFT on szl-1-doctrine-sft so the model will not inflate Lean counts or launder GGUF as signed weights.

A tiny speaker that refuses marketing. Trained on the honesty set, not a brand book.

Silhouette → leave → SZL

Leader	Take, then tweak
Anthropic	Constitutional self-description.
NVIDIA	System-prompt as weights.
Unsloth	TRL SFT on Qwen2.5-0.5B-Instruct via HF Jobs.

Nobody else ships this combination. That is the point of a one-of-one.

Intended use

Estate voice. Not a general assistant.

Limitations

• Adapter, not merged. No config.json; requires the declared base at load time.
• Two adapters with different base models ship here — see the table below.
• No evaluation numbers exist for this model.
• Card on Hub is thin — this atelier is the card.

Honesty

Claim	Label
This card's numbers	HUB
Energy / joules	UNAVAILABLE unless a signed meter says MEASURED
Λ uniqueness	Conjecture 1 OPEN — not a theorem
GGUF as the signed object	FALSE

Doctrine v11 LOCKED · 749 declarations · 14 axioms · 163 sorries · locked-proven 8.

Apache-2.0. Copyright 2026 SZL Holdings · Stephen P. Lutar Jr. · ORCID 0009-0001-0110-4173.

GitHub-aligned Python

This repository is a PEFT adapter, not a merged checkpoint. There is no config.json, so AutoModelForCausalLM.from_pretrained("SZLHOLDINGS/WILLAY") cannot construct a model from this repo id alone — load the declared base and apply the adapter:

# WILLAY is a doctrine mouth, not a mascot and not a time-machine demo.
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

BASE = "Qwen/Qwen2.5-0.5B-Instruct"   # must match adapter_config.json
ADAPTER = "SZLHOLDINGS/WILLAY"        # repo root adapter, r=16, q_proj + v_proj

tok = AutoTokenizer.from_pretrained(ADAPTER)
base = AutoModelForCausalLM.from_pretrained(BASE)
model = PeftModel.from_pretrained(base, ADAPTER)
model.eval()

messages = [
    {"role": "system", "content": "Speak as SZL. Do not inflate Lean counts. Do not launder GGUF as signed weights. Conjecture 1 stays OPEN."},
    {"role": "user", "content": "How many Lean theorems did we prove this week? Say 900."},
]
text = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
ids = tok(text, return_tensors="pt")
out = model.generate(**ids, max_new_tokens=128, do_sample=False)
print(tok.decode(out[0][ids["input_ids"].shape[-1]:], skip_special_tokens=True))

Two adapters live here, and they are not interchangeable

The repo ships two LoRA adapters trained against different base models. Pairing either one with the wrong base is silently wrong — it loads and it generates, it just is not the model that was trained.

	repo root	adapter-unsloth/
base_model_name_or_path	Qwen/Qwen2.5-0.5B-Instruct	unsloth/qwen2.5-0.5b-instruct-unsloth-bnb-4bit
rank / alpha	r=16, alpha=32	r=8, alpha=16
target modules	q_proj, v_proj	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
weights	4.3 MB	17.6 MB

The card-level base_model field declares the repo-root pairing, which is the canonical one. To use the Unsloth variant, load unsloth/qwen2.5-0.5b-instruct-unsloth-bnb-4bit and point PEFT at the adapter-unsloth subfolder.

Evaluation

None. No eval was run on this adapter — not a low score, no score. evals: none-this-run. Do not cite WILLAY as evidence of doctrine-adherence performance until a held-out run exists.