Instructions to use codercat/AtmosRegime-Qwen3-8B-SFT-RL with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use codercat/AtmosRegime-Qwen3-8B-SFT-RL with Transformers:

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

pipe = pipeline("text-generation", model="codercat/AtmosRegime-Qwen3-8B-SFT-RL")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("codercat/AtmosRegime-Qwen3-8B-SFT-RL")
model = AutoModelForCausalLM.from_pretrained("codercat/AtmosRegime-Qwen3-8B-SFT-RL", 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=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use codercat/AtmosRegime-Qwen3-8B-SFT-RL with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "codercat/AtmosRegime-Qwen3-8B-SFT-RL"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "codercat/AtmosRegime-Qwen3-8B-SFT-RL",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/codercat/AtmosRegime-Qwen3-8B-SFT-RL

SGLang

How to use codercat/AtmosRegime-Qwen3-8B-SFT-RL 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 "codercat/AtmosRegime-Qwen3-8B-SFT-RL" \
    --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": "codercat/AtmosRegime-Qwen3-8B-SFT-RL",
		"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 "codercat/AtmosRegime-Qwen3-8B-SFT-RL" \
        --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": "codercat/AtmosRegime-Qwen3-8B-SFT-RL",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use codercat/AtmosRegime-Qwen3-8B-SFT-RL with Docker Model Runner:
```
docker model run hf.co/codercat/AtmosRegime-Qwen3-8B-SFT-RL
```

AtmosRegime-Qwen3-8B-SFT-RL

Qwen3-8B fine-tuned on a few verified worked examples and then trained with reinforcement learning, from the paper Language models misapply atmospheric formulas until trained on a few worked examples. In the paper, this route repairs mechanisms that reinforcement learning alone could not: outside-regime accuracy on them rises by 61 points.

Code and data: https://github.com/acodercat/AtmosRegime. The other models from the paper are in the AtmosRegime collection.

Training

Base model: Qwen3-8B, thinking off.
Step 1, supervised fine-tuning (one epoch) on worked examples. A worked example is a program the base model wrote when the correct formula was stated in the prompt, verified by execution and paired with the problem without the formula. The mechanisms that reinforcement learning alone had not learned received 0, 1, 2 or 4 examples each, assigned at random.
Step 2, reinforcement learning: DAPO in verl with the answer reward, the same settings as the other routes (learning rate 1e-6, 128 problems × 8 samples per step, 120 steps).
Data: training fold 1 of the study's templates (29 of the 70 mechanisms); seed 42.

Use

The model was trained on one prompt format and should be given the same one. It answers with a Python program that defines solve(). Run it with thinking disabled.

# git clone https://github.com/acodercat/AtmosRegime && cd AtmosRegime   (for the prompt)
from transformers import AutoModelForCausalLM, AutoTokenizer
from atmoscoder_rl.prompts import chat

name = "codercat/AtmosRegime-Qwen3-8B-SFT-RL"
tok = AutoTokenizer.from_pretrained(name)
model = AutoModelForCausalLM.from_pretrained(name, torch_dtype="auto", device_map="auto")

problem = "Atmospheric CO2 rises from 280 ppm to 560 ppm. Calculate the radiative forcing in W m-2."
text = tok.apply_chat_template(chat(problem), tokenize=False, add_generation_prompt=True, enable_thinking=False)
inputs = tok(text, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=3072, do_sample=True, temperature=0.3)
print(tok.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

Execution and grading of the returned program are in atmoscoder_rl/evaluation/ of the repository.

Limitations

This checkpoint is one fold and one seed of a randomised design. Numbers in the paper compare mechanisms with and without examples across both folds and two seeds, so this model alone will not reproduce them.
The repair was shown on problems from the same templates. Whether it carries over to independently sourced problems of the same mechanism was not tested.
It is not a general-purpose assistant for atmospheric science.

Citation

@unpublished{ran2026regime,
  title  = {Language models misapply atmospheric formulas until trained on a few worked examples},
  author = {Ran, Maohao and Zhang, Yanting and Ma, Chendong and Song, Jun},
  year   = {2026},
  note   = {Manuscript under review}
}