AngFlow

A Creative Writing Language Model

AngFlow is a creative-writing-focused language model fine-tuned from Qwen3.5-2B-Base.

Designed to help with creative expression, storytelling, and fiction generation, AngFlow focuses on transforming writing instructions into engaging, coherent, and imaginative prose.

Rather than attempting to be a general-purpose assistant, AngFlow is specialized for creative writing tasks, with an emphasis on narrative development, descriptive language, character interactions, and storytelling.

Model Overview

Property	Details
Model Name	AngFlow
Base Model	Qwen3.5-2B-Base
Parameters	2 Billion
Model Type	Fine-tuned Language Model
Primary Specialization	Creative Writing
Architecture	Qwen3.5
Fine-tuning	Supervised Fine-Tuning (SFT)
License	Apache 2.0
Framework	Hugging Face Transformers

Creative Writing Capabilities

AngFlow is developed to assist with a range of creative-writing tasks.

1. Fiction Generation

• Short stories and long-form fiction.
• Mystery, thriller, fantasy, and science fiction.
• Contemporary and literary fiction.
• First-person and third-person narratives.

2. Narrative Development

• Story continuation and scene development.
• Plot construction and narrative progression.
• Character development and relationships.
• Conflict, tension, and suspense.

3. Descriptive Writing

• Atmospheric and immersive descriptions.
• Environmental and sensory details.
• Emotional expression and internal monologues.
• Setting and world-building.

4. Dialogue and Characterization

• Natural character conversations.
• Distinct character voices.
• Dialogue-driven storytelling.
• Character motivations and interactions.

5. Writing Instructions

• Following specified genres and writing styles.
• Maintaining narrative perspective.
• Respecting word-count constraints.
• Following tone, structure, and formatting instructions.

Intended Use

AngFlow is intended for:

• Creative writers and storytellers.
• Fiction and narrative experimentation.
• AI-powered creative-writing applications.
• Story generation and brainstorming tools.
• Developers building specialized writing assistants.
• Research into fine-tuning smaller language models for creative tasks.

Usage

Installation

pip install -U transformers accelerate torch

Generate Creative Writing

from transformers import AutoProcessor, AutoModelForMultimodalLM

MODEL_ID = "YOUR_HF_USERNAME/AngFlow"

processor = AutoProcessor.from_pretrained(MODEL_ID)

model = AutoModelForMultimodalLM.from_pretrained(
    MODEL_ID,
    torch_dtype="auto",
    device_map="auto"
)

messages = [
    {
        "role": "user",
        "content": [
            {
                "type": "text",
                "text": (
                    "Write a suspenseful short story about a night-shift "
                    "librarian who discovers that a returned book contains "
                    "a handwritten message dated twenty years in the future. "
                    "Use atmospheric descriptions and gradually build tension."
                )
            }
        ]
    }
]

inputs = processor.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=800,
    do_sample=True,
    temperature=0.85,
    top_p=0.92,
    repetition_penalty=1.08
)

generated_tokens = outputs[0][inputs["input_ids"].shape[-1]:]

response = processor.decode(
    generated_tokens,
    skip_special_tokens=True
)

print(response)

Generation Recommendations

The following are starting points for creative-writing generation, not fixed optimal settings.

Parameter	Suggested Value
Temperature	0.75–0.95
Top-p	0.85–0.95
Repetition Penalty	1.05–1.12
Max New Tokens	500–1500

Higher temperatures can encourage more varied storytelling, while lower temperatures can produce more controlled outputs.

Training Information

AngFlow is built upon the Qwen3.5-2B-Base foundation model and fine-tuned for creative-writing tasks using supervised training.

The training process focuses on teaching the model to respond to natural-language creative instructions and generate narrative-oriented responses.

Training Dataset: Curated creative-writing instruction and response examples.

Training Objective: Improve the model's ability to generate coherent, engaging, and instruction-following creative prose.

Further training configuration and dataset details can be documented as the project evolves.

Evaluation

Model quality should be assessed specifically against creative-writing benchmarks rather than relying exclusively on general language-model benchmarks.

Relevant evaluation dimensions include:

• Narrative coherence.
• Originality and creativity.
• Character consistency.
• Descriptive quality.
• Dialogue naturalness.
• Instruction adherence.
• Stylistic consistency.

Benchmark results will be added as formal evaluations are completed.

Limitations

• AngFlow is a 2-billion-parameter model and has the limitations associated with smaller language models.
• Generated stories may contain inconsistencies, repetitive passages, or predictable narrative patterns.
• Long-form storytelling may require multiple generations and human editing.
• The model can generate fictional information and should not be treated as a source of factual knowledge.
• Generated content should be reviewed for quality, originality, and suitability before publication.

Acknowledgements

AngFlow is built upon the open-source Qwen3.5 model family.

Special thanks to the Qwen team for making the underlying foundation model available under the Apache 2.0 license.

License

This project is released under the Apache License 2.0.

Please refer to the repository's license file for the complete license terms.

---

AngFlow — Built for the art of storytelling.