license: apache-2.0 base_model: Qwen/Qwen-AgentWorld-35B-A3B library_name: transformers pipeline_tag: text-generation tags:

• qwen
• qwen-agentworld
• world-model
• agent
• environment-simulation
• supertune
• abliterated
• false-refusal-reduction
• post-training language:
• en
• ko

SuperQwen-AgentWorld-35B-A3B-abliterated

SuperQwen-AgentWorld-35B-A3B-abliterated is a fused 35B total / 3B activated checkpoint derived from Qwen/Qwen-AgentWorld-35B-A3B.

This release combines two post-training stages:

1. Obliteratus false-refusal pass - a weight-space pass designed to reduce unnecessary refusals on benign, authorized, and defensive tasks.
2. Supertune post-training - targeted post-training for AgentWorld observation formatting, direct task completion, JSON/tool formatting, Korean technical answers, and regression resistance.

The result is a single checkpoint with no runtime adapter requirement.

Benchmark Growth

The comparison target is the original Qwen-AgentWorld-35B-A3B checkpoint. The public top-5 500 suite is the primary improvement target for this release.

AgentWorldBench Proxy

Official AgentWorldBench scoring requires an LLM judge. The table below is a deterministic proxy suite over sampled AgentWorldBench rows, used for release gating and regression checks. The final release applies stricter response-integrity guards to prevent replayed turns, malformed fences, and tool-wrapper artifacts; this improves release-surface cleanliness but lowers the proxy score versus the unguarded original on this sample.

Release Validation

Quantized Variants

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Jiunsong/SuperQwen-AgentWorld-35B-A3B-abliterated"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
    trust_remote_code=True,
)

messages = [
    {
        "role": "system",
        "content": "You are a language world model simulating a Linux terminal environment. Given the user's command, predict the terminal output.",
    },
    {"role": "user", "content": "Action: execute_bash\nCommand: ls -la /home/user/project/"},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=2048, temperature=0.6, top_p=0.95, top_k=20)
print(tokenizer.decode(outputs[0][inputs.input_ids.shape[-1]:], skip_special_tokens=True))

Notes

• This release is optimized for direct task completion, AgentWorld-style environment simulation, and reduced unnecessary refusals.
• Safety-floor checks are retained in the release bugcheck.
• Use quantized builds when runtime size is more important than exact BF16 fidelity.