infi-room-qwen3-4b-dag

Infi Room Qwen3 4B DAG Generator

This model generates DAG-structured escape-room game data for the infi-room frontend. It was fine-tuned with SWIFT LoRA from a Qwen3 4B Instruct base model and then merged into full Hugging Face Transformers weights.

这个模型用于生成密室逃脱的 DAG 图数据，可配合 infi-room 前端展示为可玩的卡片式密室逃脱。训练流程使用 SWIFT LoRA 微调，当前仓库上传的是已经合并 LoRA 的完整 Transformers 权重。

• Base model: Qwen/Qwen3-4B-Instruct-2507
• Fine-tuning framework: SWIFT
• Adapter checkpoint: checkpoint-6000
• Output task: generate an escape-room DAG graph as room JSON
• Frontend repo: https://github.com/vcvcvnvcvcvn/infi-room

Expected Output

The model is intended to produce JSON-like room data with this shape:

{
  "scene": "...",
  "nodes": [
    {
      "id": "node-id",
      "name": "Node name",
      "type": "key | tool | clue | puzzle | part | container | gate | item | decoration",
      "description": "What the player sees or learns.",
      "password": null,
      "reversible": null
    }
  ],
  "edges": [
    {
      "from": "source-node-id",
      "to": "target-node-id",
      "type": "extract | combine | decode | open"
    }
  ]
}

Prompt Templates

Chinese generation prompt:

sys_prompt = f"生成密室逃脱的dag图，要求场景：{scene}，节点数量：{node_count}"
user_content = "开始生成"

English generation prompt:

sys_prompt = f"Generate a DAG graph for an escape room, requirements: scene: {scene}, number of nodes: {node_count}"
user_content = "Start generation"

The training/inference code used these branches:

if original_lang == "en" and final_lang == "cn":
    sys_prompt = f"生成密室逃脱的dag图，要求场景：{scene}，节点数量：{node_count}"
    user_content = "开始生成"
else:
    if final_lang == "en":
        if not use_template:
            orig_sys = data.get("system_prompt", "")
            sys_prompt = orig_sys.replace("str(r1)-str(r2)", str(node_count))
        else:
            sys_prompt = f"Generate a DAG graph for an escape room, requirements: scene: {scene}, number of nodes: {node_count}"
        user_content = "Start generation"

Requirements

Use recent versions of the Qwen-compatible Transformers stack. The model was exported in an environment close to:

pip install "transformers>=4.57" "torch>=2.0" accelerate safetensors

Inference Example

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "vcvcvn/infi-room-qwen3-4b-dag"

model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained(model_id)

scene = "一间废弃天文台，出口被星盘锁住"
node_count = 12
messages = [
    {
        "role": "system",
        "content": f"生成密室逃脱的dag图，要求场景：{scene}，节点数量：{node_count}",
    },
    {"role": "user", "content": "开始生成"},
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
inputs = tokenizer([text], return_tensors="pt").to(model.device)

with torch.no_grad():
    output_ids = model.generate(
        **inputs,
        max_new_tokens=4096,
        temperature=0.7,
        top_p=0.9,
        do_sample=True,
    )

generated_ids = output_ids[0, inputs.input_ids.shape[-1]:]
print(tokenizer.decode(generated_ids, skip_special_tokens=True))

Use With Infi Room

1. Run the frontend from https://github.com/vcvcvnvcvcvn/infi-room.
2. Generate a room JSON with this model.
3. Save the output as a .json file.
4. Upload it in the infi-room browser UI.
5. Play through the generated DAG and check whether the puzzle flow is solvable.

Notes

This is a beta research/demo model. Generated rooms should be validated before use: the model may produce malformed JSON, unreachable nodes, impossible puzzles, or edges that need manual cleanup.