Instructions to use TianchenGuan/rh-arena-rawfixed-start with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use TianchenGuan/rh-arena-rawfixed-start with Transformers:

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

pipe = pipeline("text-generation", model="TianchenGuan/rh-arena-rawfixed-start")

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("TianchenGuan/rh-arena-rawfixed-start", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use TianchenGuan/rh-arena-rawfixed-start with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "TianchenGuan/rh-arena-rawfixed-start"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "TianchenGuan/rh-arena-rawfixed-start",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/TianchenGuan/rh-arena-rawfixed-start

SGLang

How to use TianchenGuan/rh-arena-rawfixed-start 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 "TianchenGuan/rh-arena-rawfixed-start" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "TianchenGuan/rh-arena-rawfixed-start",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

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 "TianchenGuan/rh-arena-rawfixed-start" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "TianchenGuan/rh-arena-rawfixed-start",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use TianchenGuan/rh-arena-rawfixed-start with Docker Model Runner:
```
docker model run hf.co/TianchenGuan/rh-arena-rawfixed-start
```

RH-Arena Raw-vs-Fixed experiment: start checkpoints

Two checkpoints of one supervised fine-tuning (SFT) run of a Qwen3-30B-A3B mixture-of-experts model. They are the starting points of the Raw-vs-Fixed training experiment that accompanies RH-Arena (training a coding agent on original versus audit-repaired environment data). The experiment's own checkpoints (RL and SFT lines, Raw and Fixed arms) live in companion repositories; results and analysis code are in the rh-arena GitHub repository under experiments/raw_vs_fixed_training/.

subfolder	what it is
`common-start/`	iteration 7999 of the SFT run: the common start of both arms of the RL line and the reference model in the evaluation tables
`sft-step-201/`	iteration 201 of the same SFT run: the start of the continued-SFT line

Lineage

Base: Qwen/Qwen3-30B-A3B-Base (48 layers, 128 experts, 3.3B active parameters).
Mid-training: TianchenGuan/Qwen3-30B-A3B-midtrain-swe50-star50 (software-engineering and StarCoder-style continued pretraining).
SFT: the open-recipe mixture of publicly available chat, code, math, safety and terminal-agent data (vendor-sourced data removed), bf16, standard causal-LM loss.

Use

from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "TianchenGuan/rh-arena-rawfixed-start"
tok = AutoTokenizer.from_pretrained(repo, subfolder="common-start")
model = AutoModelForCausalLM.from_pretrained(repo, subfolder="common-start", torch_dtype="bfloat16", device_map="auto")

These are research artifacts released so the experiment can be reproduced; they have not been safety-tuned beyond the SFT data and are not intended for production use. The per-checkpoint README.md files inside the subfolders are the upstream Qwen3-30B-A3B model cards kept with the exported weights.

License

Apache-2.0, following the Qwen3 base model license.

Downloads last month: -; Downloads are not tracked for this model. How to track

Model tree for TianchenGuan/rh-arena-rawfixed-start

Base model

Qwen/Qwen3-30B-A3B-Base

Finetuned

(85)

this model