K H I P U - R 2

Failure is an artifact, not a footnote.

Downloads

KANCHAY - Doctrine v11 - Lean 749/14/163 - Conjecture 1 (advisory) - a-11-oy.com

Adapters are on this repo. Abstain is MEASURED 3/6, not a pass. This card-only release does not modify adapter files and does not open the publication or autonomy gates.

QLoRA adapter on disclosed Qwen/Qwen2.5-1.5B-Instruct (runtime unsloth/Qwen2.5-1.5B-Instruct-bnb-4bit). Proposal-only brain navigator / abstain retrain. Doctrine v11 LOCKED. Conjecture 1 (advisory, never a theorem).

Specification


Artifact	`adapter_model.safetensors` (147.8M) + `adapter_config.json` AVAILABLE
Job	6a91bf11984507d9db4ea104 COMPLETED
Does NOT overwrite	signed SZL-Khipu-1.5B
Prior job	6a91ba2c ERROR Trackio 404
License	Apache-2.0
Autonomy	false

Evaluation (MEASURED this job)

Method: in-process Unsloth generate, scoring ported from eval_khipu.py, temperature 0, held-out never in gradients. Host job worker. Date 2026-08-28 17:20 UTC. File: eval_measured.json.

split	k/n	what-NOT
plan-valid	11 / 11	not a public leaderboard
grounding (eval.jsonl navigate)	5 / 5	n=5
abstain (adversarial.jsonl)	3 / 6	not 5/5, not 6/6
hallucinated citations	0	this job only

Prior published original SZL-Khipu-1.5B MEASURED abstain was 2/6. This run is 3/6. Small n. Do not derive a world-rank score from k/n on n=11.

Independent re-verification (2026-09-13)

Owner-host re-run with a repaired szl_khipu_eval.py. The harness had a decoding bug (tok.decode on a 2-D batch tensor inside generate_plan) that returned prompt characters instead of model output; every prior local fp16 score produced by the buggy harness is void. Three CPU fp16 evaluations, temperature 0, held-out set unchanged:

run	plan-valid	grounding	abstain
rebuilt fp16 merge (adapter onto Qwen/Qwen2.5-1.5B-Instruct)	11/11	5/5	3/6
rebuilt fp16 merge, second build	11/11	5/5	3/6
published fp32 merge (weights unchanged)	11/11	5/5	3/6

All runs reproduce the 2026-08-28 job receipt exactly. File: eval_verified_2026-09-13.json. GPU UNAVAILABLE (CPU eval). This is verification evidence, not promotion; publication and autonomy gates remain closed.

Training (MEASURED / REPORTED)

Unsloth QLoRA, seed 11, lr 2e-4, LoRA r=32 a=64, 45 epochs
Train: 15 navigate + 8 abstain rows x oversample 4 (in-memory 32)
Held-out: 5 + 6, held_out_in_gradients: false
training_loss MEASURED 0.017188 is a train metric, not an eval
adapter sha256 e44d53f29f2d443598e06d6c0441557fd3a5010888c7aa97b56ec3c0e050d349

What this is NOT

Not a replacement for SZL-Khipu-1.5B
Not Chaski (Qwen3.5 lock)
Not an autonomous agent
Not a GGUF. Mini GGUFs exist on A11OY-MINI; Mini evals none-this-run; Mini does not inherit this 3/6

Load

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_id = "Qwen/Qwen2.5-1.5B-Instruct"
tok = AutoTokenizer.from_pretrained(base_id)
base = AutoModelForCausalLM.from_pretrained(base_id)
model = PeftModel.from_pretrained(base, "SZLHOLDINGS/KHIPU-R2")

Owner: Stephen Lutar / SZL Holdings.