Text Generation
PEFT
Safetensors
English
qlora
governed-agent
proposal-only
research-only
szl-holdings
khipu
abstain-retrain
conversational
Instructions to use SZLHOLDINGS/KHIPU-R2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use SZLHOLDINGS/KHIPU-R2 with PEFT:
from peft import PeftModel from transformers import AutoModelForCausalLM base_model = AutoModelForCausalLM.from_pretrained("unsloth/Qwen2.5-1.5B-Instruct-bnb-4bit") model = PeftModel.from_pretrained(base_model, "SZLHOLDINGS/KHIPU-R2") - Notebooks
- Google Colab
- Kaggle
Disclose KHIPU-R2 abstain job; never overwrite signed 1.5B
Browse files
README.md
CHANGED
|
@@ -10,6 +10,15 @@ tags:
|
|
| 10 |
- szl-holdings
|
| 11 |
- khipu
|
| 12 |
- series-a
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 13 |
---
|
| 14 |
|
| 15 |
# KHIPU-R2 — cutting
|
|
@@ -19,14 +28,18 @@ Fall 2026 collection. **No weights in this repository yet.** The card is the loo
|
|
| 19 |
**Silhouette:** small instruct (Qwen3 / Gemma 3 class).
|
| 20 |
**Our cut:** receipt-native agent. It drafts. It refuses. It never executes. A controller outside the weights validates JSON, requires human approval, and mints a DSSE receipt only after approved action.
|
| 21 |
|
|
|
|
|
|
|
| 22 |
| | |
|
| 23 |
|---|---|
|
| 24 |
| **House** | [SZL Holdings](https://huggingface.co/SZLHOLDINGS) |
|
| 25 |
-
| **Line** | Khipu
|
| 26 |
-
| **
|
|
|
|
| 27 |
| **License** | Apache-2.0 |
|
| 28 |
| **Λ** | Conjecture 1 (never a theorem) |
|
| 29 |
| **Doctrine** | v11 LOCKED · 749 / 14 / 163 |
|
|
|
|
| 30 |
|
| 31 |
## What it will do
|
| 32 |
|
|
@@ -40,12 +53,13 @@ Fall 2026 collection. **No weights in this repository yet.** The card is the loo
|
|
| 40 |
- Not an authorizer or executor.
|
| 41 |
- Not a factual oracle.
|
| 42 |
- Not a copy of another house's weights. Fine-tune on SZL-owned, rights-cleared curriculum only.
|
|
|
|
| 43 |
|
| 44 |
## Lineage (intended)
|
| 45 |
|
| 46 |
Previous ready-to-wear: [`SZL-Khipu-1.5B`](https://huggingface.co/SZLHOLDINGS/SZL-Khipu-1.5B) (Qwen2.5-1.5B QLoRA) and [`szl-receiptagent-qwen35-0.8b-v2`](https://huggingface.co/SZLHOLDINGS/szl-receiptagent-qwen35-0.8b-v2) (MEASURED 5/5 JSON + 6/6 adversarial refusals on a small preregistered gate, not a public benchmark).
|
| 47 |
|
| 48 |
-
R2 is the recut: newer small-instruct silhouette, same house rules, card-before-weights.
|
| 49 |
|
| 50 |
## Evidence bound
|
| 51 |
|
|
|
|
| 10 |
- szl-holdings
|
| 11 |
- khipu
|
| 12 |
- series-a
|
| 13 |
+
szl:
|
| 14 |
+
doctrine: v11-LOCKED
|
| 15 |
+
artifact_class: ADAPTER
|
| 16 |
+
originality: FINETUNE_DISCLOSED_BASE
|
| 17 |
+
line: KHIPU-R2
|
| 18 |
+
jobs: RUNNING
|
| 19 |
+
job_id: 6a91ba2c45686a1580c12020
|
| 20 |
+
job_namespace: SZLHOLDINGS
|
| 21 |
+
never_overwrite: SZLHOLDINGS/SZL-Khipu-1.5B
|
| 22 |
---
|
| 23 |
|
| 24 |
# KHIPU-R2 — cutting
|
|
|
|
| 28 |
**Silhouette:** small instruct (Qwen3 / Gemma 3 class).
|
| 29 |
**Our cut:** receipt-native agent. It drafts. It refuses. It never executes. A controller outside the weights validates JSON, requires human approval, and mints a DSSE receipt only after approved action.
|
| 30 |
|
| 31 |
+
This repo is the **abstain retrain** target. It does **not** overwrite signed [`SZL-Khipu-1.5B`](https://huggingface.co/SZLHOLDINGS/SZL-Khipu-1.5B).
|
| 32 |
+
|
| 33 |
| | |
|
| 34 |
|---|---|
|
| 35 |
| **House** | [SZL Holdings](https://huggingface.co/SZLHOLDINGS) |
|
| 36 |
+
| **Line** | Khipu successor. Signed 1.5B stays immutable. |
|
| 37 |
+
| **HF Jobs** | Abstain retrain **RUNNING** [`6a91ba2c45686a1580c12020`](https://huggingface.co/jobs/SZLHOLDINGS/6a91ba2c45686a1580c12020). Push target **this repo**. Not MEASURED. Trackio URL UNKNOWN. |
|
| 38 |
+
| **Status** | CUTTING · ROADMAP until a receipt lands |
|
| 39 |
| **License** | Apache-2.0 |
|
| 40 |
| **Λ** | Conjecture 1 (never a theorem) |
|
| 41 |
| **Doctrine** | v11 LOCKED · 749 / 14 / 163 |
|
| 42 |
+
| **Serve** | Do not load this into the Khipu lab. Lab serves signed GGUF only. |
|
| 43 |
|
| 44 |
## What it will do
|
| 45 |
|
|
|
|
| 53 |
- Not an authorizer or executor.
|
| 54 |
- Not a factual oracle.
|
| 55 |
- Not a copy of another house's weights. Fine-tune on SZL-owned, rights-cleared curriculum only.
|
| 56 |
+
- Not a write to `SZL-Khipu-1.5B`. A mis-aimed `SZL-Khipu-1.5B-abstain` job was cancelled.
|
| 57 |
|
| 58 |
## Lineage (intended)
|
| 59 |
|
| 60 |
Previous ready-to-wear: [`SZL-Khipu-1.5B`](https://huggingface.co/SZLHOLDINGS/SZL-Khipu-1.5B) (Qwen2.5-1.5B QLoRA) and [`szl-receiptagent-qwen35-0.8b-v2`](https://huggingface.co/SZLHOLDINGS/szl-receiptagent-qwen35-0.8b-v2) (MEASURED 5/5 JSON + 6/6 adversarial refusals on a small preregistered gate, not a public benchmark).
|
| 61 |
|
| 62 |
+
R2 is the recut: newer small-instruct silhouette, same house rules, card-before-weights. Evals none this run until finish + receipt.
|
| 63 |
|
| 64 |
## Evidence bound
|
| 65 |
|