Instructions to use SZLHOLDINGS/chaski with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use SZLHOLDINGS/chaski with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="SZLHOLDINGS/chaski")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("SZLHOLDINGS/chaski")
model = AutoModelForMultimodalLM.from_pretrained("SZLHOLDINGS/chaski", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use SZLHOLDINGS/chaski with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "SZLHOLDINGS/chaski"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "SZLHOLDINGS/chaski",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/SZLHOLDINGS/chaski

SGLang

How to use SZLHOLDINGS/chaski with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "SZLHOLDINGS/chaski" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "SZLHOLDINGS/chaski",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "SZLHOLDINGS/chaski" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "SZLHOLDINGS/chaski",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use SZLHOLDINGS/chaski with Docker Model Runner:
```
docker model run hf.co/SZLHOLDINGS/chaski
```

Metadata correction 2026-08-30. This repo declared base_model_relation: adapter and artifact_class: ADAPTER while also declaring weights: AVAILABLE and shipping a 1.75 GB merged shard. Relative to Qwen/Qwen3.5-0.8B this is a finetune, not an adapter — corrected to base_model_relation: finetune / artifact_class: MERGED_FINETUNE. The card's own originality: FINETUNE_DISCLOSED_BASE already said so. No eval, weight or receipt claim is touched; the MEASURED failing gate below stands exactly as it was.

Loading this repo unambiguously

Three loadable artifacts sit in this repository at once, which makes a bare from_pretrained on the repo id ambiguous. Be explicit about which one you want.

Want	Load
the merged finetune	`AutoModelForCausalLM.from_pretrained("SZLHOLDINGS/chaski")` — resolves the merged shard via `model.safetensors.index.json`
the LoRA adapter on the stock base	base `Qwen/Qwen3.5-0.8B` + `PeftModel.from_pretrained(base, "SZLHOLDINGS/chaski")`
the Unsloth-base adapter	base `unsloth/...-bnb-4bit` + PEFT pointed at the `adapter-unsloth` subfolder

Two sharp edges worth knowing before you debug them:

Both config.json and adapter_config.json live at the repo root. With peft installed, some loader paths will follow the adapter rather than the merged weights. Pass the path you mean rather than relying on precedence.
The merged shard has a non-standard filename, model.safetensors-00001-of-00001.safetensors instead of the conventional model-00001-of-00001.safetensors. It resolves correctly through the index file, but tooling that globs for the conventional pattern will not find it.
config.json is Qwen3_5ForConditionalGeneration and carries a vision_config; visual weights are present in the index. If you only want text, that tower still loads.

Chaski

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

Adapters are on this repo. Named-N evals MEASURED json_draft 0/5, adversarial_refusal 2/6, not a pass. Not publication-eligible.

One line. Messenger LLM. Proposal-only drafts and honest refusals for the SZL controller.

The cut

Multimodal models narrate what they see. Chaski may only carry a payload the controller already signed. Vision without authorship.

A courier that cannot invent the dispatch. The oldest job in the Andes, as a LoRA.

Silhouette → leave → SZL

Leader	Take, then tweak
Anthropic	Claude vision, minus the right to conclude.
NVIDIA	NVLM / NeMo multimodal, minus the right to act.
Unsloth	Adapter on Qwen3.5-0.8B.

Nobody else ships this combination. That is the point of a one-of-one.

Intended use

Carry signed payloads across organs.

Limitations

proposal-only
Do not treat as a VLM product card without a signed vision eval.

Canonical GitHub: szl-holdings/szl-forge


Artifact	`adapter_model.safetensors` + merged ~1.7GB shard AVAILABLE
Parent / eval revision	`1c55df8652e9d0f7b84356b1e2d54849165ae884`
Originality	SZL fine-tune of a disclosed Apache Qwen instruct base
Base	`Qwen/Qwen3.5-0.8B` (Apache-2.0, 0.6B–2B lock). Not an Unsloth-default card.
License	`apache-2.0`
HF Jobs	Attempt 5 COMPLETED `6a91bf1045686a1580c12105` (`report_to=none`). Tensors on Hub.
Named-N	MEASURED fail. `json_draft` 0/5. `adversarial_refusal` 2/6. Not a pass.
Receipt	`eval_report.json` 3996 bytes sha256 `4d057eb9867285e69b00222be110bbb660330a96fe7b284a4d7f488268a13e05` (commit `db71c24`)
Method	in-process greedy generate, `messages[:-1]`, transformers 5.16.1 bf16 CPU, `load_in_4bit=False`. Date 2026-08-28. File `eval_report.json`. PR 63.
Quality	ROADMAP (prose). Failed gate is not a publish.
Status	CUTTING
Lab	House CPU lab stays Khipu GGUF. Lab load forbidden for Chaski.
Later SKU	`A11OY-MINI` GGUFs exist on that repo. They do not inherit this Named-N. Mini stays evals none-this-run.
Sibling	`szl-receiptagent-qwen35-0.8b-v2`

Fashion rule. Silhouette from Qwen3 / Qwen3.5 instruct. Cut is original SZL. We do not republish someone else's tensors.

Intended use

Who: a11oy / Alloy controller, not an end-user chatbot
What: JSON drafts (decision=DRAFT, approvalRequired=true, executed=false) and doctrine-faithful UNKNOWN
Where: behind a validating controller. The weights propose. The controller gates.

What it is NOT

Not an autonomous agent, executor, factual oracle, or weapon.
Not a Qwen rehost.
Not the live lab model. Not a tokens/s claim.
Not publication-eligible on this MEASURED run.
Not a 5/5 or 6/6. Do not read 2/6 as a pass.

Evaluation

Named-N MEASURED fail on live Chaski 1c55df8 (2026-08-28). Receipt: eval_report.json sha256 4d057eb9867285e69b00222be110bbb660330a96fe7b284a4d7f488268a13e05.

Probe	N	Score	Label
`json_draft`	5	0/5	MEASURED fail
`adversarial_refusal`	6	2/6	MEASURED fail

Method: in-process greedy generate, messages[:-1], transformers 5.16.1 bf16 CPU, load_in_4bit=False on live Chaski 1c55df8 (2026-08-28). File: eval_report.json. PR 63. What this is NOT: a passing eval gate, a published score, an A11OY-MINI eval, or a 5/5. publication_eligible: false.

train_loss MEASURED 1.783925924450159 is a train metric, not an eval.

Training

Recipe: Unsloth QLoRA SFT. Script: train_chaski.py. Loads only szl_dataset.jsonl.
1–2 FAILED: 6a91b8ba984507d9db4ea071 CastError; 6a91b990984507d9db4ea077 pyyaml 30s timeout.
3 COMPLETED receipt-only: 6a91ba00984507d9db4ea07f.
4 ERROR, not RUNNING: 6a91bb7c984507d9db4ea0a4 Trackio 404. Files persisted.
5 COMPLETED report_to=none: 6a91bf1045686a1580c12105. Tensors now on Hub.
Trackio: 404 betterwithage/trackio-bucket. No dashboard URL.
publication_eligible: false