Chitrapathak V3 DocRecon 5K β€” LoRA Adapter

A PEFT/LoRA adapter for document reconstruction on document-page images, fine-tuned from the Chitrapathak v3 lineage on a 5,000-sample mixed dataset (task13_5k). The model takes a rendered document page and reconstructs its content structure β€” text blocks, reading flow, and document formatting.

Status: experimental research artifact. Not production-evaluated. An A/B evaluation against the prior production adapter had not been run at training completion (2026-11-16).

Model lineage

krutrim-ai-labs/Chitrapathak-2                        (base VLM)
  └── chitrapathak_v3_targeted_structure_task12       (2026-11-16)
        └── chitrapathak_v3_document_reconstruction_5k (this adapter, 2026-11-16)
  • Base model: krutrim-ai-labs/Chitrapathak-2 (server snapshot e21a8b29)
  • Adapter size: ~239 MB (LoRA, adapter_model.safetensors)
  • SHA256: 89dddd86eeee876ffb65d59d963cc4ce32565b1631a18210df59e76c3eb981d6

Intended use

  • Document-page reconstruction for Indic-language government forms and scanned documents (Marathi, Gujarati, English mixed corpus)
  • Research evaluation of document-reconstruction capability on top of the v3 targeted-structure checkpoint

Training details

Setting Value
Steps 1,760 (2 epochs, seed 3407)
Learning rate 5e-06
Batch micro-batch 1/GPU Γ— 6 GPUs (DDP), grad-accum 1
Precision bf16 + gradient checkpointing
Max image side 1,280 px
Max target tokens 4,096
Train set task13_5k/datasets/train_mix_final.jsonl (β‰ˆ5k samples)
Val set task13_5k/datasets/val_task12.jsonl

Usage

from peft import PeftModel

# 1. load the base model exactly as Chitrapathak-2 documents it
base_model = ...  # e.g. AutoModelForCausalLM / vision model per base model card

# 2. attach this adapter
model = PeftModel.from_pretrained(base_model, "snxtyle/chitrapathak-v3-docrecon-5k")

Processor, tokenizer, and chat template files are bundled in this repo (processor_config.json, tokenizer.json, tokenizer_config.json, chat_template.jinja) so the adapter directory is load-ready alongside the base model.

Limitations & caveats

  • No published benchmark results yet β€” evaluate before relying on it.
  • Doc-reconstruction runs inherit the base model's Indic OCR weaknesses; Gujarati script fidelity historically lags Marathi in this lineage.
  • Trained with max_side=1280 β€” pages rendered much larger may shift in quality.
  • The previous A/B in this lineage (chitrapathak_v3_marguj_gov_candidate_20261112, a different run) scored BASELINE_BETTER; treat this adapter as a candidate, not a promotion.

License

Follows the base model (krutrim-ai-labs/Chitrapathak-2); verify its license terms before any redistribution or production use.

Provenance

Trained on internal infrastructure (6Γ— H100, DDP). Full training config (training_config.json), logs, and per-epoch checkpoints are retained in the internal experiments directory chitrapathak_v3_document_reconstruction_5k_20261117/.

Downloads last month
9
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for snxtyle/chitrapathak-mgb-v1

Adapter
(4)
this model