SZLHOLDINGS

KHIPU-R2

README

License: apache-2.0

The cut

Leaders ship v1 and changelog the rest. We name the retraining of silence as its own model. Failure is an artifact, not a footnote.

A public retrain whose only job is to improve one metric: honest abstain under adversarial handles.

Silhouette → leave → SZL

Table with columns: Leader, Take, then tweak
Leader	Take, then tweak
Anthropic	Red-team → constitution update. We red-team → adapter.
NVIDIA	Recipe re-run with a new seed and a bounded delta.
Unsloth	Same FastLanguageModel loop, new curriculum, new evidence record.

This card makes the combination inspectable: a separate abstain-retraining SKU, bounded small-sample results, and an exact-source documentation publisher that cannot modify weights or promotion state.

Intended use

Continue the abstain-retrain loop. Keep publication and autonomy gates closed until independently verifiable evaluation evidence satisfies the governed release criteria.

Limitations

research-only
No signed R2 eval receipt in this atelier.
Publishing this card is a documentation update, not model promotion or autonomy approval.

Canonical GitHub: szl-holdings/szl-forge/khipu_r2

Specification

Table

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

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.

Table with columns: split, k/n, what-NOT
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.

Training (MEASURED / REPORTED)

Unsloth QLoRA, seed 11, lr 2e-4, LoRA r=32 α=64, 45 epochs
Train: 15 navigate + 8 abstain rows × 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

python
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.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

SZLHOLDINGS

Model Tree

Base

Qwen/Qwen2.5-1.5B-Instruct

Adapter

this model

Input Modalities

Text

Output Modalities