| --- |
| license: apache-2.0 |
| task_categories: |
| - text-generation |
| language: |
| - en |
| tags: |
| - szl-holdings |
| - unsloth |
| - qlora |
| - receipts |
| - doctrine-v11 |
| - proposal-only |
| pretty_name: Receipted Unsloth (SZL method) |
| --- |
| |
| # Receipted Unsloth |
|
|
| How SZL Holdings actually trains. Silhouette from Unsloth QLoRA. Cut is original SZL. We do not republish Unsloth Studio, Desktop, copy, code, or someone else's tensors. |
|
|
| Collection: [Receipted Unsloth — LIVE](https://huggingface.co/collections/SZLHOLDINGS/receipted-unsloth-live-6a91baafa72e0b7a20f4f7d5) |
|
|
| ## The house loop |
|
|
| 1. Disclose the Apache base (`Qwen/Qwen2.5-*` or `Qwen/Qwen3.5-0.8B`). |
| 2. Train with Unsloth `FastLanguageModel` QLoRA on owner metal or HF Jobs (`uv run` + `HF_TOKEN`). |
| 3. Bind dataset SHA-256, LoRA knobs, seed, and loss into a training receipt. |
| 4. Push adapter or merge only with that receipt. No receipt = ROADMAP, not a model. |
| 5. Serve GGUF on `szl-model-inference-lab`. `szl-forge-lab` is a SNAPSHOT evidence console, not a trainer. |
| 6. Weights propose (`decision=DRAFT`, `approvalRequired=true`, `executed=false`). A controller outside the weights gates. A human approves. |
|
|
| ## Repeated knobs (MEASURED in scripts, not a performance claim) |
|
|
| | Knob | House value | |
| |---|---| |
| | seed | 11 | |
| | LoRA r / alpha | 16 / 32 | |
| | lr | 2e-4 | |
| | optim | adamw_8bit | |
| | loss | response-only CE (Chaski) | |
| | doctrine rows | 41 MEASURED in `SZLHOLDINGS/szl-1-doctrine-sft` | |
| | Λ | Conjecture 1, never a theorem | |
| | doctrine | v11 LOCKED 749/14/163 | |
| |
| Source kit: [szl-holdings/szl-forge](https://github.com/szl-holdings/szl-forge) (`train_szl.py`, Unsloth QLoRA, GGUF rebirth). Hub Jobs scripts: `SZLHOLDINGS/chaski/train_chaski.py`. |
| |
| ## LIVE vs CUTTING (2026-08-28) |
| |
| | Artifact | Class | Evidence | |
| |---|---|---| |
| | [SZL-Forge-1.5B-ReceiptAgent](https://huggingface.co/SZLHOLDINGS/SZL-Forge-1.5B-ReceiptAgent) | LIVE weights | owner-signed train/eval receipts | |
| | [SZL-Khipu-1.5B](https://huggingface.co/SZLHOLDINGS/SZL-Khipu-1.5B) | LIVE weights | trained 2026-07-14 on `betterwithage`, loss 0.0245 **REPORTED** | |
| | [SZL-Khipu-1.5B-GGUF](https://huggingface.co/SZLHOLDINGS/SZL-Khipu-1.5B-GGUF) | LIVE derivative | served on inference-lab, sha256 matched | |
| | [szl-receiptagent-qwen35-0.8b-v2](https://huggingface.co/SZLHOLDINGS/szl-receiptagent-qwen35-0.8b-v2) | LIVE adapter | Unsloth tag; 5/5 + 6/6 on a tiny owner gate **MEASURED** for that gate only | |
| | [chaski](https://huggingface.co/SZLHOLDINGS/chaski) | CUTTING | Jobs retry `6a91ba00` RUNNING on a10g-small. Adapter not uploaded. No eval this run. | |
| | KHIPU-R2 / A11OY-MINI / WILLAY Hub model | CUTTING / ROADMAP | cards, not weights | |
| |
| HF Jobs exist under org `SZLHOLDINGS`. A job id is not MEASURED training. Trackio on the personal free account returned HTTP 402; logs fell back local. Next runs should pin Trackio to the org Team plan, not `@betterwithage`. |
| |
| ## What we refuse |
| |
| - Unsloth Studio / Desktop UI, Tauri wrapper, dual-dialect `/v1/messages` clone. |
| - Relabeling Khipu GGUF as Chaski. |
| - Fake tokens/s. ReceiptAgent 28 tok / 16.274s stays on that card only. |
| - Bricklayer / claude.ai / NIM console chrome. |
| - Invented evals. `evals: none-this-run` until a held-out k/n exists. |
| |
| ## Famous on purpose |
| |
| Hub visitors should land on **receipted weights they can run**, not 30 empty cards. Pin this collection. Pin [killinchu](https://huggingface.co/spaces/SZLHOLDINGS/killinchu) and [a11oy](https://huggingface.co/spaces/SZLHOLDINGS/a11oy) as the two shipping flagships. When Chaski finishes with a receipt, add it here. Until then it stays CUTTING. |
| |
