PranavKeshav's picture
Update README.md
ac7315c verified
metadata
license: apache-2.0
datasets:
  - ciol-research/global-festivals-wiki
  - WorkWithData/cities
  - Yelp/yelp_review_full
  - dair-ai/emotion
language:
  - en
  - kn
  - hi
  - es
  - zh
  - te
  - de
  - ko
  - sq
  - fr
  - id
  - pl
  - it
  - vi
  - tr
  - ru
  - he
  - ar
  - fa
  - bn
  - th
  - ja
base_model:
  - google/gemma-3-1b-it
pipeline_tag: text-generation

Gemma-3-1B Event-Planner (4-bit QLoRA)

Adapter-only repo for a culturally sensitive event-planning assistant fine-tuned via LoRA on google/gemma-3-1b-it.
This adapter (~50 MB) can be applied to the 4-bit base model at inference time, so you don’t need to ship multi-GB merged weights.

Base model: google/gemma-3-4b-it
Fine-tuned with: LoRA r=8, α=32, dropout=0.05, 4-bit NF4 quant.

Intended use

Generates culturally sensitive event plans (weddings, baby-naming, college fests …).
Asks clarifying questions about culture, guest count, budget, dietary needs.

Training data

  • dair-ai/emotion (3 k / 0.5 k)
  • ciol-research/global-festivals-wiki (9 k / 1 k)
  • corbt/all-recipes (15 k / 1.5 k)
  • WorkWithData/cities (6 k / 1 k)
  • Yelp/yelp_review_full (12 k / 2 k)

Model Details

  • Base model: google/gemma-3-1b-it (4 B parameters, instruction-tuned)
  • Quantization: 4-bit NF4 via bitsandbytes
  • LoRA config:
    • rank r = 8
    • α = 32
    • dropout = 0.05
    • target modules = ["q_proj","v_proj"]
  • Trainable params: ~0.75 M (0.07 % of base)

Fine-tuning data (≈ 75 k examples total)

Domain Dataset Train / Val Why included
Emotion & Tone dair-ai/emotion 3 k / 0.5 k Adapt style & follow-up questioning
Cultural Festivals ciol-research/global-festivals-wiki 9 k / 1 k Rituals, symbols, dates across cultures
Cuisine & Menus corbt/all-recipes 15 k / 1.5 k Authentic recipes for menu planning
Venue / Geodata WorkWithData/cities 6 k / 1 k Real cities + coords for location tips
Vendors & Services Yelp/yelp_review_full 12 k / 2 k Business vocabulary & recommendation tone

Local Usage

# Install runtime dependencies:
pip install accelerate==1.7.0 bitsandbytes==0.45.5 peft==0.15.2 sentencepiece==0.2.0 torch==2.7.0 transformers==4.51.3 trl==0.17.0

# Load the 4-bit base + adapter:
import torch
from transformers import (
    AutoTokenizer,
    AutoModelForCausalLM,
    BitsAndBytesConfig,
    pipeline,
)
from peft import PeftModel

# 1. Quant config
bnb_cfg = BitsAndBytesConfig(
    load_in_4bit              = True,
    bnb_4bit_quant_type       = "nf4",
    bnb_4bit_use_double_quant = True,
    bnb_4bit_compute_dtype    = torch.float16,
)

# 2. Tokenizer
BASE = "google/gemma-3-1b-it"
tokenizer = AutoTokenizer.from_pretrained(
    BASE,
    trust_remote_code=True,
    use_auth_token=True,
)

# 3. Base model
base = AutoModelForCausalLM.from_pretrained(
    BASE,
    quantization_config=bnb_cfg,
    device_map="auto",
    torch_dtype=torch.float16,
    trust_remote_code=True,
    use_auth_token=True,
)

# 4. LoRA adapter
ADAPTER = "PranavKeshav/event-planner-gemma-4bit"
model   = PeftModel.from_pretrained(
    base,
    ADAPTER,
    device_map="auto",
    torch_dtype=torch.float16,
    use_auth_token=True,
)
model.eval()

# 5. Pipeline
pipe = pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
    device_map="auto",
    max_new_tokens=150,
    temperature=0.7,
    top_p=0.9,
)

# 6. Test
print(pipe("Plan a Gujarati wedding for 120 guests in Ahmedabad.")[0]["generated_text"])

Model Card & Citation

  1. Intended use: Generate culturally sensitive event plans; ask clarifying questions about dates, budgets, dietary needs.
  2. Limitations: May hallucinate or miss rare cultural details; verify all critical recommendations.
  3. License: Same as google/gemma-3-1b-it (Apache-2.0) + dataset licenses; see individual datasets.
  4. Citation:
    @misc{gemma_event_planner_2025,
       title = {Gemma-3-4B Event-Planner LoRA Adapter},
       author = {Keshav, Pranav},
       year = {2025},
       howpublished = {\url{https://huggingface.co/YOUR_USERNAME/event-planner-gemma-4bit}}
     }