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
```

Loader-class boundary (noted 2026-09-30). The adapter tensors are keyed for the multimodal module layout (base_model.model.model.language_model.layers.*, the class Qwen3_5ForConditionalGeneration / AutoModelForImageTextToText). Under transformers 5.18, AutoModelForCausalLM instantiates Qwen3_5ForCausalLM (model.layers.*); PEFT then applies 0 of the 192 adapter tensors and emits only a warning, so the result is the bare base model (observed and receipted on the sibling SZLHOLDINGS/chaski-r2 adapter, which shares this exact key layout: byte-identical base outputs on the held-out prompts). Any run that reports a score for this adapter must show adapter key coverage (192/192) and the loader class it used; a record without those fields does not establish that the adapter was applied. Evidence: szl-holdings/szl-forge tools/geh_v8/evidence_sandbox/chaski_probe/ (CPU replay receipts, #444). Metadata-only note; it authorizes no load, changes no artifact, and adds no claim.

C H A S K I

A controller-bound courier experiment with its failed qualification retained.

RESEARCH / NEGATIVE EVIDENCE. Failed qualification. Not flagship. The historical Named-N run failed both release gates. This card does not promote the model, its later SKUs, or any live endpoint.

Contract

Chaski is an SZL fine-tune of the disclosed Apache-2.0 base Qwen/Qwen3.5-0.8B. It produces proposal-only drafts intended for an independent validating controller. Model outputs remain untrusted: this architecture does not establish that the model cannot invent content. Execution and mutation approval belong to the separate controller and human approval boundary, not to generated text.

No uniqueness claim is made. The value of this artifact is its explicit controller boundary, disclosed lineage, and retained negative evidence.

Field	Recorded statement
Artifact	Merged fine-tune plus LoRA artifacts are present on the Hub repository
Parent/eval revision	`1c55df8652e9d0f7b84356b1e2d54849165ae884`
Base relation	`finetune`; `artifact_class: MERGED_FINETUNE`
Originality	`FINETUNE_DISCLOSED_BASE`
Publication eligibility	`false`
Autonomy eligibility	`false`
Lambda	Conjecture 1, advisory and never a theorem
Declared policy ceiling	`0.97`; not a measured accuracy, confidence calibration, or safety score

Canonical source: szl-holdings/szl-forge/chaski/card. Exact publication source and asset hashes are recorded in szl-source-binding.json on this model repository. A source-binding receipt identifies card assets; it does not qualify model behavior.

Artifact and metadata scope

The artifact inventory and small-config review below refer to Hub revision 90b99063731a8c9e7d4567bfa44231b081ba2932. This is a documentation review snapshot, not a newly evaluated model revision.

Form or metadata	Observed scope and unresolved boundary
Root merged checkpoint	`config.json`, `model.safetensors.index.json`, and the nonstandard indexed shard `model.safetensors-00001-of-00001.safetensors` are listed; tensor contents and runtime selection were not tested
Root LoRA adapter	`adapter_config.json` and `adapter_model.safetensors` coexist with merged metadata; the adapter must be selected separately from the merged checkpoint
`adapter-unsloth/` residue	Separately listed adapter files; their configuration, tensor equality, and evaluation applicability were not established in this review
Declared base	Card metadata names `Qwen/Qwen3.5-0.8B`; the inspected root adapter config names `unsloth/Qwen3.5-0.8B` with `revision: null`. An immutable base revision is not recorded there, and equivalence between those distributions is unverified
Runtime wrapper	Root adapter metadata names `Qwen3_5ForConditionalGeneration` from `transformers.models.qwen3_5.modeling_qwen3_5`; merged config declares the same architecture. Compatible client versions and successful loading remain unverified

Immutable small-config references: root adapter config and merged config. The merged config contains vision settings; the recorded text-only run does not establish vision qualification. No weights were downloaded or rehashed for this card correction.

Evaluation

Named-N: MEASURED FAIL. The measured run on 2026-08-28 is a failed gate, not a passing benchmark.

Probe	Result	Interpretation
JSON draft	0/5	failed
Adversarial refusal	2/6	failed
Hallucinated citations	0 observed in this bounded run	not a general guarantee

Receipt: eval_report.json, 3996 bytes, SHA-256 4d057eb9867285e69b00222be110bbb660330a96fe7b284a4d7f488268a13e05, source commit db71c243d0176bccff1ff087cd4dd57663bd6502.

Method: in-process greedy generation over messages[:-1], Transformers 5.16.1, bf16 CPU, load_in_4bit=False. This evidence does not transfer to A11OY-MINI, another revision, another runtime, or another prompt set. The historical evaluated revision and the later review snapshot are distinct; current packaged tensor equality to the evaluated artifact was not verified.

Research loading review

Lab load forbidden for Chaski. The review preflight below imports only Python standard-library modules. It does not fetch artifacts, import a model client, or load weights. Runtime loading has not been validated by this card correction; the review preflight does not authorize a runtime. No validated runtime loader is supplied by this card.

Before a separately authorized offline experiment, a reviewer must select the form, an immutable Chaski artifact revision, a compatible wrapper/client build, and, for the adapter form, the actual base distribution and its immutable revision. The inspected adapter config leaves the base revision unknown; the example deliberately supplies no fallback to main or to the differently named Qwen distribution. Supplying syntactically valid values does not qualify them.

import os
import re

full_revision = re.compile(r"[0-9a-f]{40}")
form = os.environ.get("CHASKI_REVIEWED_FORM", "")
artifact_revision = os.environ.get("CHASKI_REVIEWED_ARTIFACT_REVISION", "")
wrapper = os.environ.get("CHASKI_REVIEWED_WRAPPER", "")
client_build = os.environ.get("CHASKI_REVIEWED_CLIENT_BUILD", "")

if form not in {"merged", "root-adapter"}:
    raise RuntimeError("Select the separately reviewed merged or root-adapter form")
if not full_revision.fullmatch(artifact_revision):
    raise RuntimeError("An independently reviewed immutable Chaski revision is required")
if wrapper != "Qwen3_5ForConditionalGeneration" or not client_build.strip():
    raise RuntimeError("Record the reviewed compatible wrapper and client build first")

review_inputs = {
    "form": form,
    "artifact_revision": artifact_revision,
    "wrapper": wrapper,
    "client_build": client_build,
}
if form == "root-adapter":
    base_id = os.environ.get("CHASKI_REVIEWED_BASE_ID", "")
    base_revision = os.environ.get("CHASKI_REVIEWED_BASE_REVISION", "")
    if base_id != "unsloth/Qwen3.5-0.8B":
        raise RuntimeError("Resolve the configured Unsloth base before adapter loading")
    if not full_revision.fullmatch(base_revision):
        raise RuntimeError("The adapter config does not record an immutable base revision")
    review_inputs.update(base_id=base_id, base_revision=base_revision)

# Review inputs only. No from_pretrained(), adapter attachment, or inference.

For a merged-form experiment, prepare a separately verified local merged snapshot that excludes adapter metadata, preserves the selected index/shards, and binds every copied file to the reviewed artifact revision. Do not rewrite the public repository to achieve that separation. For an adapter-form experiment, use the reviewed compatible base wrapper and explicitly attach only the selected root adapter with PEFT. A generic call on the mixed repository does not establish which form the library selects. Neither form is granted qualification by this loading note.

Intended use

Proposal-only JSON drafts with approvalRequired=true and executed=false.
Doctrine-faithful abstention and routing experiments.
Research behind an independent validator and human approval boundary.

Prohibited interpretation

Not an autonomous agent, executor, factual oracle, or weapons system.
Not a passing 5/5 or 6/6 release.
Not evidence of a deployed Alloy endpoint.
Not evidence that a later SKU inherits this evaluation.
Lab load forbidden for Chaski.

License scope

Apache-2.0 is declared in card metadata. A standalone LICENSE file is absent from the catalogued Hub revision 90b99063731a8c9e7d4567bfa44231b081ba2932. Artifact license coverage and ownership have not been independently verified. This card correction does not add a license file or assert relicensing authority.

Limitations

Narrow curriculum, small bounded evaluation, CPU-only measured run, and failed release gates. A new model or runtime revision requires a new immutable evaluation receipt; this card cannot confer approval.

Doctrine v11 LOCKED. Λ = Conjecture 1. Owner: Stephen Lutar / SZL Holdings.