chitrapathak_marguj_mixed_v2

LoRA continuation of snxtyle/chitrapathak_marguj_mixed for Marathi / Gujarati OCR + HTML table structure, trained on lighton_marguj_unified_v3.

What this is

• Base model: krutrim-ai-labs/Chitrapathak-2
• Init adapter: snxtyle/chitrapathak_marguj_mixed (continued, not trained from scratch)
• Adapter type: PEFT LoRA (r=32, α=64, dropout=0.05)
• Data: LightOn English tables (50%) + Indic table/mixed (30%) + Indic normal OCR (~20%)
• Train recipe: 2 epochs, lr 5e-6, batch 1, grad accum 8, max_side 1280, seed 3407
• Local experiment: chitrapathak_lighton_marguj_unified_v3_20261101

Load

from peft import PeftModel
from transformers import AutoModelForImageTextToText, AutoProcessor

base = "krutrim-ai-labs/Chitrapathak-2"
adapter = "snxtyle/chitrapathak_marguj_mixed_v2"

model = AutoModelForImageTextToText.from_pretrained(base)
model = PeftModel.from_pretrained(model, adapter)
processor = AutoProcessor.from_pretrained(base)

Notes

• Strong gain on HTML table detection / complete-table rate vs chitrapathak_marguj_mixed.
• Mild regression possible on pure Marathi normal OCR; keep v1 mixed as rollback for prose-only SLAs.
• Serve with the canonical train prompt; no system message; no repetition penalty; max image side 1280.