Instructions to use SZLHOLDINGS/chaski-r2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use SZLHOLDINGS/chaski-r2 with PEFT:

from peft import PeftModel
from transformers import AutoModelForCausalLM

base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-0.8B")
model = PeftModel.from_pretrained(base_model, "SZLHOLDINGS/chaski-r2")

Transformers

How to use SZLHOLDINGS/chaski-r2 with Transformers:

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

pipe = pipeline("text-generation", model="SZLHOLDINGS/chaski-r2")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("SZLHOLDINGS/chaski-r2")
model = AutoModelForCausalLM.from_pretrained("SZLHOLDINGS/chaski-r2", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.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(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use SZLHOLDINGS/chaski-r2 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "SZLHOLDINGS/chaski-r2"
# 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-r2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/SZLHOLDINGS/chaski-r2

SGLang

How to use SZLHOLDINGS/chaski-r2 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-r2" \
    --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-r2",
		"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-r2" \
        --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-r2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Unsloth Desktop
Docker Model Runner
How to use SZLHOLDINGS/chaski-r2 with Docker Model Runner:
```
docker model run hf.co/SZLHOLDINGS/chaski-r2
```

SEPARATE RESEARCH ADAPTER · MEASURED GATE RECORD · NAMED-N ARTIFACT BINDING INCOMPLETE · LATER GATE FAIL · NOT PROMOTABLE

Chaski-R2

What this is

A separate research and proposal-only SKU based on Qwen/Qwen3.5-0.8B. The root training_receipt.json describes bf16 LoRA with r=16 and alpha=32. It does not overwrite SZLHOLDINGS/chaski and is separate from SZLHOLDINGS/chaski-5050.

Evidence and artifact identity

Reviewed Hub revision: 4ef29684c56de07626b1cefb4b93a1b486de2aa1.

Published record	Scope and receipt-reported result
eval_named_n_generate.json	Owner-run named-N record, `label=MEASURED`, `gate_ran=true`, JSON drafts 5/5, adversarial refusals 6/6, `publication_eligible=false`; computed 2026-08-29.
eval_merged_measured.json	Merged checkpoint, plain Transformers fp16, temperature 0, named PR 63 gates: JSON drafts 5/5 and adversarial refusals 6/6; evaluated `2026-09-14T12:34:01.477237+00:00`.
merge_receipt.json	Repair on 2026-09-14 after the previous merge was an empty base copy. Root adapter applied to that base copy; 96/96 modules touched; maximum weight delta `0.00423942506313324`.

The separate post-publication gate receipt names Hub revision 4ef29684c56de07626b1cefb4b93a1b486de2aa1 and reports 1/5 on the canonical held-out suite in transformers-cpu mode, with overall FAIL, heldout_passed=false, productionAuthorization=false, and publicationAuthorization=false. This is a different suite from the earlier named-N record; its results cannot be exchanged or combined with those earlier scores. The receipt reports refusalPassed=true for the separate refusal check while its embedded held-out object has refusal_no_regression=false. Those fields are preserved as reported; neither changes its overall FAIL result.

What the evidence does and does not establish

The 5/5 + 6/6 records are real published measured records. Their public receipt files do not cryptographically bind those results to one exact public adapter-file SHA-256 or merged-weight revision. The separate later failed gate explicitly names the Hub revision above; this does not retroactively supply the missing binding for the earlier successful records. No third-party benchmark, broad capability estimate, or qualification follows.

The older szl-source-binding.json retains evaluation_state=NONE_THIS_RUN and release_blocker=no_json_or_refusal_gate; the training_receipt.json and training_receipt.local.json retain evals=none-this-run. These historical fields conflict with a present-tense claim that no evaluation records exist. They are preserved as source history, not silently rewritten.

Training receipts also describe different runs: training_receipt.json reports 10 rows and train loss 1.4571; training_receipt.local.json reports 32 rows and train loss 0.7656. Train loss is not evaluation. The adapter-unsloth/training_receipt.json describes a separate subdirectory adapter with r=8 and alpha=16; evidence for the root adapter must not be transferred to it. No single historical adapter digest is asserted here as the current public file identity. The provenance.json retains blocked lineage, consent, privacy-review, deployment, and served-revision boundaries.

Adapter-only evaluation binding (2026-09-25)

The 2026-09-25 evaluation binding is adapter-only: it binds the 2026-09-24 named-fixture bake-off run of the root adapter to adapter_model.safetensors SHA-256 6f12981ea5df5e22d3493eefb20d75db75c1c88961ba6621a35af09edd0cdde6 (with adapter_config.json SHA-256 d36472a3100231dfdf0c4aa16d3a80cb8ab53c86af33a15172b06f9d9a313193) at Hub revision 661f8ee9ff6ab8b11dda6e7a9d42c14d3124c6dd, and to the published LF source bytes of its receipt and identity manifest recorded in szl-holdings/szl-forge@3a2ad95a. It does not bind the merged model.safetensors, the separate adapter-unsloth/ adapter, or the A11OY-MINI R2 GGUF, and it does not retroactively bind the earlier named-N records above. That run reports JSON draft contracts 5/5 and refusal prefixes 6/6 for the adapter on the declared Qwen base (base comparator 0/5 and 6/6) on small reused named fixtures; this is not a held-out generalization or broad quality result. The later post-publication gate FAIL remains in force, production disposition remains HOLD, and this evidence is UNSIGNED_HONEST. Exact copies of the three records are in evidence/2026-09-25-adapter-binding/ (source_publication_binding.json, receipt.json, identity.json).

Release, autonomy, and deployment boundary

publication_eligible=false and autonomy_eligible=false remain unchanged. Proposal only. No qualification, promotion, deployment authorization, autonomy, house-lab load, or lab approval follows from these records. No tokens-per-second or energy-performance claim is made. Lambda uniqueness remains Conjecture 1: open and advisory, never a theorem.

Relationship and source of truth

Chaski-R2 is a sibling of Chaski and Chaski-5050. The A11OY-MINI R2 GGUF is a separate derived runtime with its own evidence boundary; results do not transfer between adapters, merged checkpoints, and quantizations.

Use the immutable evidence links above and the canonical post-publication gate receipt for evaluation claims. The historical source card is chaski-r2/card/README.md; its none-this-run language and szl-source-binding.json document a prior publication state, not a current complete evaluation summary.