ganesh-gemma4-e4b-v3

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

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 🙏