Instructions to use baseten/caesar-cipher-sgd-qwen38_27b-expert with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use baseten/caesar-cipher-sgd-qwen38_27b-expert with Transformers:

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

pipe = pipeline("image-text-to-text", model="baseten/caesar-cipher-sgd-qwen38_27b-expert")
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)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("baseten/caesar-cipher-sgd-qwen38_27b-expert")
model = AutoModelForMultimodalLM.from_pretrained("baseten/caesar-cipher-sgd-qwen38_27b-expert", 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=256)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use baseten/caesar-cipher-sgd-qwen38_27b-expert with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "baseten/caesar-cipher-sgd-qwen38_27b-expert"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "baseten/caesar-cipher-sgd-qwen38_27b-expert",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/baseten/caesar-cipher-sgd-qwen38_27b-expert

SGLang

How to use baseten/caesar-cipher-sgd-qwen38_27b-expert 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 "baseten/caesar-cipher-sgd-qwen38_27b-expert" \
    --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": "baseten/caesar-cipher-sgd-qwen38_27b-expert",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

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 "baseten/caesar-cipher-sgd-qwen38_27b-expert" \
        --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": "baseten/caesar-cipher-sgd-qwen38_27b-expert",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use baseten/caesar-cipher-sgd-qwen38_27b-expert with Docker Model Runner:
```
docker model run hf.co/baseten/caesar-cipher-sgd-qwen38_27b-expert
```

Caesar Cipher — Qwen3.8-27B SGD expert

This repository contains the final post-trained inference checkpoint at optimizer update 100 from full-weight GRPO with SGD, trained to decode Caesar-cipher text. The completed fresh trajectory consumed 6,400 training responses.

Base: Qwen/Qwen3.8-27B, pinned revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0.
Environment: Caesar Cipher from Reasoning Gym, with a frozen task bank and completed-answer binary reward adapter. This is not an official upstream benchmark result.
Training: training run rktht87u; full provenance.

Checkpoint	Accuracy	Correct / scored	Truncated
Own pinned-base baseline	40.25%	161 / 400	2 / 400
Final update 100	92.25%	369 / 400	10 / 400

The observed gain is 52.00 percentage points. Each evaluation is 50 held-out tasks × 8 sampled completions. Accuracy is mean binary correctness, including truncation failures in the denominator; it is not best-of-8 or pass-at-8. Final accuracy is 92.25%; the earlier validation 90 score of 93% is preserved as an intermediate result and is not the published endpoint.

Setting	Value
Optimizer	SGD, constant LR `0.1`
Momentum / weight decay	`0.0` / `0.0`
Gradient clip norm	Not explicitly overridden; native default not asserted
Policy clip lower / upper	`0.2` / `0.28`
KL / entropy coefficient	`0.0` / `0.0`
Training / validation bank	114 / 50 tasks
Batch / optimizer budget	8 questions × 8 completions; 100 updates
Precision	BF16
Thinking	Enabled, medium effort
Sampling	Temperature 1.0, top-p 0.95, top-k 20
Output / context tokens	16,384 / 32,768
Consumption	Ordered async batches, maximum completed-update lag 1
Validation / checkpoint interval	Every 10 updates

Use the supplied tokenizer/chat template with thinking enabled and reasoning_effort="medium" for comparable text generation. The original runtime pins Transformers 5.12.1. The evaluation bank was reused through training; no separate final confirmation set, replicated optimizer-causality study, cross-task transfer, or multimodal evaluation was performed.

Final native 100, the 400-completion final evaluation, durable final inference/full-state checkpoints, completed platform archive and GPU release are verified. All 11 original validation summaries (updates 0–100) have been recovered and verified. Full raw-trace export and coverage remain unverified; the complete validation history does not imply a completed raw-trace audit.

The source is miles-rg-caesar-cipher-sgd100-20261022-bf673f5b467f/rg-caesar-cipher-sgd100-20261022-bf673f5b467f/training/eval-snapshots/step_99 from job qrzx40w / project qvvnjxq. Its original final evaluation timestamp is 2026-10-23T06:34:49.547636+00:00. The original training run is Finished, but its summary remains at update 90. Original training history and all scheduled native validation summaries have since been recovered for the plots below. This card uses original native final 100, not that stale summary. Original graded outcomes and binary rewards remain unchanged.

Only final post-trained inference weights and required inference assets are included. Repacking must preserve every tensor bit: no cast, quantization, averaging or base-weight substitution. Publication requires a separate source/output tensor-hash proof, exact 18-shard layout, and remote-file verification; this card alone does not certify transfer completion. Resumable model/optimizer/FP32-master/scheduler/RNG/cursor checkpoints and raw traces remain in Baseten archives and owning experiment evidence.

Apache-2.0, following the pinned base model; see LICENSE and the pinned base-model card.

Training and validation curves

The plots below show the recorded training trajectory and verified validation results for this environment. Training uses a running average of 10 with the original series visible; validation is unsmoothed. No training, evaluation, or interpolation was performed to create these images.

Training metric: rollout/raw_reward against rollout/step. Validation metric: eval/validation against completed optimizer updates.

Both published experts contain 100 optimizer updates. Their training curves use the full native axis 0–99; validation includes every scheduled point from 0 through 100. Restored runs retain their original checkpoint ancestry and exclude abandoned branches.

The overlay compares these recorded trajectories; it does not establish an optimizer-only causal effect. See each model’s settings and limitations above.