README.md

---
license: gemma
base_model: google/gemma-4-E4B-it
tags: [ganesha, marathi, hindi, sanskrit, on-device, litert, lora, devotional]
language: [mr, hi, en, sa]
pipeline_tag: text-generation
---

# Ganesh SLM — Gemma 4 E4B fine-tune

An offline assistant fine-tuned on Shri Ganesha's shlokas, stotras, aartis, rituals and stories,
for Ganeshotsav 2026. Answers in Marathi, Hindi or English. **No RAG** — the canon is in the weights.

**Created in service of Shri Ganesh by Ravi Kadam — https://www.linkedin.com/in/ravikadam/**

## What it does

Recites, verbatim and exactly:

- **Sukhkarta Dukhharta** (Samarth Ramdas)
- **Ganapati Atharvashirsha** — complete, sections 1–19 plus the closing shanti
- **Sankatnashan Ganesh Stotra** (Narada Purana)
- **Ganapati Ashtottara Shatanamavali** — all 108 names
- **Ganesha Pancharatnam** (Adi Shankara) — all five verses
- **Vakratunda Mahakaya**, **Ganapati Gayatri**

Also covers puja vidhi (pranapratishtha, shodashopachara, durva and the 21 patri, uttarpuja,
visarjan), the Puranic stories, the Ashtavinayak, Mumbai's mandals, and 2026 festival dates.

## What it deliberately refuses

- **Sankashti moonrise times** — vary by city and month; it defers to a panchang
- **Any year other than 2026** — its calendar is frozen at 2026 and it says so
- **Live queue or darshan timings** — it is offline and cannot know

Every dated answer is **year-stamped** ("In 2026...") so it never reads as a claim about the current
year, and every muhurat names its city (Mumbai's 2026 window is 11:20–13:48, about 18 minutes later
than the generic figure — which is exactly why a bare time is wrong).

## Known limitations — please read

- **It can fabricate.** Asked for a stotra it does not know (e.g. "Ganesha Vajrakavacha"), it may
  recite a different stotra instead of declining. Do not trust an unfamiliar text from it.
- **Four texts are NOT included** because they could not be verified against a reliable source:
  Ghalin Lotangan, Mantrapushpanjali, Shendur Lal Chadhayo, and the mool mantras. Ghalin Lotangan
  normally follows Sukhkarta Dukhharta in the aarti sequence — this model does not know it.
- Canonical texts were verified against public sources, **not against a printed pothi**. Editions
  differ in punctuation, anusvara/conjunct conventions, and section numbering.
- It is **not a guruji and not a panchang**. For muhurat, family vidhi and anything disputed, ask
  your elders and your priest.

## Method

Small curated corpus (32 units) → deterministic pair generation. Canonical Devanagari is spliced
**byte-exact** from YAML and never passes through a generative model; only the *question* side is
varied. Unverified texts are gated out of training entirely. 1,959 training pairs
(1,216 verbatim / 405 calendar / 338 prose).

LoRA r=64, alpha=128, 3 epochs, lr 1e-4, bf16, seq 2048, on an L40S.
Adapters scoped to the language model — Gemma 4's E-series wraps vision/audio projections in
`Gemma4ClippableLinear`, which PEFT cannot target.

Final train loss 0.46, train token accuracy 0.97.

## Usage

```python
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

m_id = "ravikadam/ganesh-gemma4-e4b"
tok = AutoTokenizer.from_pretrained(m_id)
model = AutoModelForCausalLM.from_pretrained(m_id, dtype=torch.bfloat16, device_map="auto")

msgs = [
  {"role": "system", "content": "You are an offline assistant fine-tuned on Shri Ganesha's shlokas, stotras, aartis, rituals and stories. Answer in the language the user writes in. Never invent a verse."},
  {"role": "user", "content": "सुखकर्ता दुखहर्ता आरती म्हण."},
]
enc = tok.apply_chat_template(msgs, add_generation_prompt=True, tokenize=True,
                              return_dict=True, return_tensors="pt").to(model.device)
out = model.generate(**enc, max_new_tokens=700, do_sample=False)
print(tok.decode(out[0][enc["input_ids"].shape[-1]:], skip_special_tokens=True))
```

गणपती बाप्पा मोरया 🙏