Gemma-3-1b-korean-novel

license: gemma language:

• ko
• en pipeline_tag: text-generation tags:
• korean
• gemma
• roleplay
• comment-generation
• diary

---

Gemma-3-1B Korean Novel Commenter

A lightweight conversational model designed to write comments on users’ diary or SNS-style posts.

Instead of general chatting, the model is specialized for reaction writing: reading a personal text and replying as if a real person left a comment under the post.

This model is used in the mobile app: https://play.google.com/store/apps/details?id=com.chensi.offgram&hl=ko

Provided Files

Currently recommended:

• comment_270m_260204.gguf — very fast, low resource inference
• comment_1b_260203.gguf — higher quality responses

The 270M version is intended for on-device usage. The 1B version produces more natural emotional reactions but still keeps low compute requirements.

---

What the Model Does

The model reads a diary entry or SNS-like post and generates a human-like reaction comment.

Typical outputs include:

• empathy
• encouragement
• emotional reaction
• casual conversation-style replies

It is not a general knowledge assistant and will perform poorly on factual Q&A tasks.

---

Prompt Format

System

너는 SNS 일기에 댓글을 다는 사람이다.

User

[COMMENTER_PROFILE]$ageString ${character.gender}
[COMMENTER_STYLE]${character.speechStyle}
[COMMENT_LENGTH]${character.reactionLength}
[LANG]$lang
[DIARY]$diaryContent
[COMMENT]

---

Prompt Field Explanation

COMMENTER_PROFILE

• Social age group expression (e.g., early 20s, mid-30s, late-40s)
• Gender: male / female

COMMENTER_STYLE Defines personality and speaking style based on ai_comment.json.

Examples:

• polite and considerate
• warm empathy
• casual teenage friendliness
• calm encouragement

COMMENT_LENGTH Controls approximate response length. Values typically: 1, 1~2, 1~3 (roughly number of sentences)

LANG Final output language. Currently supported:

• ko
• en (more planned)

DIARY Actual user diary text written inside the app.

---

Training Characteristics

• Base: Gemma family small model
• Task: emotional reaction generation
• Data: primarily empathy-focused conversational text
• Optimized for: supportive responses rather than diverse knowledge

Because of the small model size (especially 270M), responses may be simple and sometimes repetitive. The model prioritizes emotional appropriateness over linguistic variety.

Continuous updates are planned to improve diversity and naturalness.

---

Limitations

• Limited factual accuracy
• Repetition in long conversations
• Weak reasoning capability
• Not suitable for coding or academic Q&A

This model is intended for social interaction simulation, not an information assistant.

---

Intended Use

• diary apps
• SNS companion systems
• roleplay interaction
• emotional feedback systems
• character comment generation

---

Notes

The model is designed to respond kindly and supportively to personal writing. It may over-empathize or avoid negative reactions due to training bias.

---

License

gemma

---

Citation

If you use this model in research or applications, please reference the Hugging Face repository.