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

Libraries

How to use SlowGuess/ABForge-Qwen3-8B-RL with Transformers:

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

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

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("SlowGuess/ABForge-Qwen3-8B-RL")
model = AutoModelForCausalLM.from_pretrained("SlowGuess/ABForge-Qwen3-8B-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=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use SlowGuess/ABForge-Qwen3-8B-RL with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "SlowGuess/ABForge-Qwen3-8B-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": "SlowGuess/ABForge-Qwen3-8B-RL",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/SlowGuess/ABForge-Qwen3-8B-RL

SGLang

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

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

ABForge-Qwen3-8B-RL

The RL-only ablation of the unified ABForge model: rubric-guided GRPO applied directly to Qwen/Qwen3-8B, with no SFT warm start (update 200). This is the "RL only" row of the paper's post-training ablation.

Renamed on 2026-08-13 from ABForge-Qwen3-8B-Combined-RL; the weights are unchanged.

Tasks

Both ABForge tasks in one checkpoint — Task 1: Ablation Objective Identification (propose Target Module / Research Question pairs for a paper) and Task 2: Ablation Experiment Synthesis (produce an executable ablation plan).

Training

Mixed-task GRPO on train/RL_task1_30K.jsonl + train/RL_task2_30K.jsonl from SlowGuess/abforge-data, a 1:1 task mixture, with each rollout routed to its task-specific reward by data_source. Unlike the released model, this run starts from the base model rather than from ABForge-Qwen3-8B-SFT.

Results

AblationBench, automated rubric-based LLM-as-a-Judge evaluation (eval/ablationbench_200.jsonl, 200 papers, judge claude-sonnet-4-6):

Stage	Task 1	Task 2
`Qwen/Qwen3-8B` (base)	44.4	43.4
`-SFT` (unified, SFT only)	30.7	52.2
This model (unified, RL only)	52.2	54.9
`ABForge-Qwen3-8B` (unified, SFT → GRPO)	55.9	62.4

RL alone already lifts both tasks over the base model, but initializing joint RL from the mixed SFT checkpoint is worth a further +3.7 on Task 1 and +7.5 on Task 2 — i.e. SFT is an effective RL initialization even though it does not improve Task 1 on its own.

Per-paper generations and judge rationales are released in the dataset repo under outputs/task{1,2}/{generations,judge_claude-sonnet-4-6}/abforge-rl.jsonl, and the aggregate row is abforge-rl in outputs/leaderboard.csv.