dungnvt/qwen35-2b-general

Model Card for qwen35-2b-general (OpenClaw General/Story Expert)

This model is a fine-tuned version of unsloth/Qwen3.5-2B. It is specifically trained as the General/Story Expert for the OpenClaw Mixture of Experts (MoE) architecture.

It has been trained using TRL and Unsloth.

Model Details

Model Description

The qwen35-2b-general model is designed to handle everyday questions, write short stories, and help with simple daily tasks (like drafting a message or rewriting text) where no tools are needed. It acts as a friendly, general assistant.

• —Developed by: OpenClaw Project
• —Model type: Causal Language Model (MoE Expert)
• —Language(s) (NLP): English, Vietnamese
• —License: Apache 2.0
• —Finetuned from model: unsloth/Qwen3.5-2B

Intended Uses & Limitations

This expert model is intended to be used as a component within the OpenClaw MoE router system. Its primary role is to process requests that:

• —Require natural and clear answers for everyday questions without using external tools.
• —Involve creative writing, such as short stories in a simple, engaging style.
• —Entail simple daily tasks (drafting messages, rewriting text).

Negative Prompts (What it should NOT do):

• —Decide the best tool workflow before acting.
• —Summarize multiple fetched documents into a market analysis.

Training Details

Training Data

The model was fine-tuned on the expert_general_story.jsonl dataset consisting of 200 high-quality synthetic examples formatted in the standard Qwen/HuggingFace ChatML format.

Training Procedure

The training was performed using the Supervised Fine-Tuning (SFT) configuration with TRL and Unsloth for optimization.

Quick start

import torch
from unsloth import FastLanguageModel

max_seq_length = 4096

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "checkpoints/qwen35-2b-general",
    max_seq_length = max_seq_length,
    dtype = torch.float32,
)

FastLanguageModel.for_inference(model)

user_question = "Viết giúp tôi một câu chuyện ngắn về tình bạn."
prompt = f"<|im_start|>user\n{user_question}<|im_end|>\n<|im_start|>assistant\n<think>\n"

inputs = tokenizer(
    text = prompt,
    return_tensors = "pt",
    add_special_tokens = False
).to("cuda")

outputs = model.generate(
    **inputs, 
    max_new_tokens = 1024,
    use_cache = True,
    pad_token_id = tokenizer.eos_token_id, 
    repetition_penalty = 1.15,
    temperature = 0.6,
    top_p = 0.95,
    top_k = 20,
    min_p = 0.00,
    do_sample = True           
)

response = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]
answer = response.split("assistant\n")[-1].strip()

print(answer)

Routing Configuration (Mergekit)

For integration into the OpenClaw MoE via mergekit-moe, the following positive prompts are recommended for routing:

• —"Answer naturally and clearly for everyday questions when no tools are needed."
• —"Write a short story in a simple engaging style with a clear beginning, middle, and end."
• —"Help with a simple daily task such as drafting a message or rewriting text."
• —"Giải thích ngắn gọn"
• —"Viết giúp tôi"
• —"Kể một câu chuyện ngắn"