Chaski-5050

Adapters are on this repo. Evals none-this-run. Not MEASURED.

Owner-GPU recut on an RTX 5050 Laptop (bf16 LoRA, not QLoRA). Original SZL cut of disclosed Apache Qwen/Qwen3.5-0.8B. Not a republish of Qwen tensors. Not an Unsloth-default card. CUTTING.

Artifact	adapter_model.safetensors 25,587,104 bytes AVAILABLE (sha256 620b3488fac2ebc6518090424de5b3c6a182293cf52dfd5bd9f886f54aef0df5)
Job	local-5050 (owner metal, not an HF Job)
Does NOT overwrite	SZLHOLDINGS/chaski
Dataset	SZLHOLDINGS/szl-1-doctrine-sft · 41 rows · jsonl sha256 ddc5594b…0611a243
Publication / autonomy	false / false
License	Apache-2.0

Evaluation

Status: none-this-run. No JSON/refusal gate ran. Not 5/5. Not 6/6. Do not load this ID into the Khipu lab.

train_loss MEASURED 2.228136855544466 is a train metric, not an eval (method: Unsloth trainer log, N=41 rows, 3 epochs, 33 steps, train_runtime 883.2224s, 2026-08-28 17:56 UTC). File: training_receipt.json.

Training (MEASURED this run)

• Recipe: train_chaski_bf16_5050.py · Unsloth 2026.7.2 · transformers 5.5.0 · torch 2.10.0+cu128
• GPU: NVIDIA GeForce RTX 5050 Laptop, 7.96 GB
• LoRA r=16 α=16, bf16, batch 1, ga 4, lr 2e-4, adamw_8bit, seed 11, max_seq 2048
• copied_live_chaski_weights: false

What this is NOT

• Not live Chaski
• Not a Qwen rehost
• Not a GGUF (A11OY-MINI stays ROADMAP)
• Not production. Lab load forbidden.

Load

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_id = "Qwen/Qwen3.5-0.8B"
tok = AutoTokenizer.from_pretrained(base_id)
base = AutoModelForCausalLM.from_pretrained(base_id)
model = PeftModel.from_pretrained(base, "SZLHOLDINGS/chaski-5050")

Doctrine v11 LOCKED. Λ = Conjecture 1 (advisory, never a theorem). Owner: Stephen Lutar / SZL Holdings.