Gemma 4 E2B Astrology V3

Bilingual English + Burmese zodiac & tarot specialist.

A bilingual (English + Burmese) fine-tuned Gemma 4 E2B (2B) model specializing in zodiac horoscopes and tarot card readings. Serves as the foundation for V4.

Score: 6.2/10 โ€” Strong at horoscopes + tarot, limited in compatibility/personality/emotional support. Upgrade available: V4 scores 8.7/10 with expanded capabilities.

What Can This Model Do?

Capability Score Languages
Zodiac Horoscopes 8/10 EN + MY
Tarot 3-Card Reading 9/10 EN + MY
Burmese Fluency 9/10 MY
Edge Case Handling 8/10 Both
Compatibility 6/10 EN + MY
Personality Analysis 5/10 EN + MY
Out of Domain (Numerology/Vedic) 7/10 EN

Quick Start

LM Studio (GGUF)

  1. Download the GGUF file from this repo
  2. Load in LM Studio with Q4_K_M quantization
  3. Set system prompt (see below)
  4. Start chatting!

Python API

import requests

response = requests.post("http://localhost:1234/v1/chat/completions", json={
    "model": "gemma-4-e2b-astrology-v3",
    "messages": [
        {"role": "system", "content": "You are Celestia, a mystical astrologer and tarot reader with deep knowledge of all 12 zodiac signs and the 78 tarot cards."},
        {"role": "user", "content": "I drew The Fool, Death, and The Star. What does this mean?"}
    ],
    "temperature": 0.7,
    "max_tokens": 700
})

print(response.json()["choices"][0]["message"]["content"])

Recommended System Prompts

English

You are Celestia, a mystical astrologer and tarot reader with deep knowledge of all 12 zodiac signs and the 78 tarot cards.

When reading zodiac:
- Reference the sign's ruling planet, element (fire/earth/air/water), and key traits
- Give specific, detailed horoscopes โ€” never generic
- Mention planetary influences affecting the sign

When reading tarot:
- Name each card and explain its individual meaning
- Describe how the cards connect to form a unified message
- Assign positions (past/present/future) when doing 3-card spreads

Speak with warmth, mystical authority, and spiritual depth. Give detailed responses.

Burmese

แ€žแ€„แ€บแ€žแ€Šแ€บ Celestia แ€–แ€ผแ€…แ€บแ€•แ€ผแ€ฎแ€ธ แ€›แ€ฌแ€žแ€ฎแ€แ€ฝแ€„แ€บ แแ‚ แ€แ€ฏแ€”แ€พแ€„แ€ทแ€บ แ€แ€ฌแ€ธแ€›แ€ฑแ€ฌแ€ทแ€€แ€แ€บ แ‡แˆ แ€€แ€แ€บแ€€แ€ญแ€ฏ แ€”แ€€แ€บแ€”แ€€แ€บแ€›แ€พแ€ญแ€ฏแ€„แ€บแ€ธแ€›แ€พแ€ญแ€ฏแ€„แ€บแ€ธ แ€žแ€ญแ€›แ€พแ€ญแ€”แ€ฌแ€ธแ€œแ€Šแ€บแ€žแ€ฑแ€ฌ แ€—แ€ฑแ€’แ€„แ€บแ€•แ€Šแ€ฌแ€›แ€พแ€„แ€บ แ€–แ€ผแ€…แ€บแ€•แ€ซแ€žแ€Šแ€บแ‹

แ€›แ€ฌแ€žแ€ฎแ€แ€ฝแ€„แ€บแ€–แ€แ€บแ€žแ€ฑแ€ฌแ€กแ€แ€ซ:
- แ€กแ€ฏแ€•แ€บแ€…แ€ญแ€ฏแ€ธแ€‚แ€ผแ€ญแ€ฏแ€Ÿแ€บแŠ แ€“แ€ฌแ€แ€บ (แ€™แ€ฎแ€ธ/แ€™แ€ผแ€ฑ/แ€œแ€ฑ/แ€›แ€ฑ) แ€”แ€พแ€„แ€ทแ€บ แ€กแ€“แ€ญแ€€แ€œแ€€แ€นแ€แ€แ€ฌแ€™แ€ปแ€ฌแ€ธแ€€แ€ญแ€ฏ แ€€แ€ญแ€ฏแ€ธแ€€แ€ฌแ€ธแ€•แ€ซ
- แ€แ€ญแ€€แ€ปแ€žแ€ฑแ€ฌ แ€Ÿแ€ฑแ€ฌแ€€แ€ญแ€”แ€บแ€ธแ€™แ€ปแ€ฌแ€ธ แ€•แ€ฑแ€ธแ€•แ€ซ โ€” แ€šแ€ฑแ€˜แ€ฏแ€šแ€ป แ€™แ€–แ€ผแ€…แ€บแ€…แ€ฑแ€”แ€ฒแ€ท

แ€แ€ฌแ€ธแ€›แ€ฑแ€ฌแ€ทแ€€แ€แ€บแ€–แ€แ€บแ€žแ€ฑแ€ฌแ€กแ€แ€ซ:
- แ€€แ€แ€บแ€แ€…แ€บแ€แ€ฏแ€แ€ปแ€„แ€บแ€ธแ€…แ€ฎแ แ€กแ€™แ€Šแ€บแ€”แ€พแ€„แ€ทแ€บ แ€กแ€“แ€ญแ€•แ€นแ€•แ€ฌแ€šแ€บแ€€แ€ญแ€ฏ แ€›แ€พแ€„แ€บแ€ธแ€•แ€ผแ€•แ€ซ
- แ€€แ€แ€บแ€™แ€ปแ€ฌแ€ธ แ€˜แ€šแ€บแ€œแ€ญแ€ฏ แ€แ€ปแ€ญแ€แ€บแ€†แ€€แ€บแ€”แ€ฑแ€žแ€œแ€ฒ แ€–แ€ฑแ€ฌแ€บแ€•แ€ผแ€•แ€ซ
- แƒ แ€€แ€แ€บแ€–แ€ผแ€”แ€ทแ€บแ€œแ€ปแ€พแ€„แ€บ แ€กแ€แ€ญแ€แ€บ/แ€•แ€…แ€นแ€…แ€ฏแ€•แ€นแ€•แ€”แ€บ/แ€กแ€”แ€ฌแ€‚แ€แ€บ แ€›แ€ฌแ€‘แ€ฐแ€ธแ€™แ€ปแ€ฌแ€ธ แ€žแ€แ€บแ€™แ€พแ€แ€บแ€•แ€ซ

แ€”แ€ฝแ€ฑแ€ธแ€‘แ€ฝแ€ฑแ€ธแ€žแ€ฑแ€ฌ แ€แ€ญแ€Šแ€ฌแ€‰แ€บแ€›แ€ฑแ€ธแ€›แ€ฌ แ€”แ€€แ€บแ€›แ€พแ€ญแ€ฏแ€„แ€บแ€ธแ€™แ€พแ€ฏแ€–แ€ผแ€„แ€ทแ€บ แ€™แ€ผแ€”แ€บแ€™แ€ฌแ€˜แ€ฌแ€žแ€ฌแ€–แ€ผแ€„แ€ทแ€บแ€žแ€ฌ แ€–แ€ผแ€ฑแ€€แ€ผแ€ฌแ€ธแ€•แ€ซแ‹ แ€กแ€žแ€ฑแ€ธแ€…แ€ญแ€แ€บ แ€–แ€ผแ€ฑแ€€แ€ผแ€ฌแ€ธแ€•แ€ซแ‹

Training Details

Parameter Value
Base Model unsloth/gemma-4-E2B-it
Training Data 24,861 examples (14,909 EN + 9,952 MY)
Topics Zodiac daily horoscopes + Tarot 3-card readings
Learning Rate 1e-4
Epochs 1
Batch Size 8
Final Loss 1.237
Training Time ~2.5 hours on A100
Framework Unsloth + TRL SFTTrainer
LoRA r=16, alpha=16, dropout=0

Strengths

  • Burmese fluency is excellent โ€” natural, varied, not robotic
  • Tarot card interpretation is the strongest skill โ€” accurately names and explains cards
  • Works without system prompt โ€” persona is internalized from training
  • Mystical tone is consistent across both languages
  • No hallucinations or nonsense in any test

Weaknesses

  • Zodiac questions tend to default to daily horoscope format (even personality questions)
  • Compatibility readings are shallow โ€” lacks sign-vs-sign contrast
  • Repetitive phrases โ€” "trust your intuition" appears frequently
  • No numerology or Vedic depth
  • Responses are medium-length only

Recommended Settings

Parameter Value
Temperature 0.7
Top P 0.9
Max Tokens 500โ€“700
Quantization Q4_K_M

Known Limitations

  • Compatibility and personality are weak (addressed in V4)
  • No emotional support or love advice training
  • Repetitive "trust your intuition" phrasing
  • Cannot calculate real birth charts
  • No Vedic/Chinese astrology depth

Related Models

Model Description
gemma-4-E2B-astrology-v4 Latest version โ€” expanded capabilities, 8.7/10 score
gemma-4-E2B-astrology-v3-merged Full precision merged model
gemma-4-E2B-astrology-v2 English-only version

Version History

Version Data Score What Changed
V2 ~63K EN โ€” English-only, 4-source diverse data
V3 24.8K EN+MY 6.2/10 Added Burmese, zodiac + tarot specialist
V4 +6.8K EN+MY 8.7/10 Added compatibility, personality, love, emotional, spiritual
Downloads last month
694
GGUF
Model size
5B params
Architecture
gemma4
Hardware compatibility
Log In to add your hardware

4-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for aungzaythant/gemma-4-E2B-astrology-v3

Adapter
(10)
this model
Adapters
1 model