Sanskrit-Qwen2.5-VL-7B-Instruct-OCR

This model is a fine-tuned version of Qwen/Qwen2.5-VL-7B-Instruct on the diabolic6045/sanskrit-ocr-parallel-corpus-chat-template dataset. This data is converted from snskrt/Sanskrit_OCR_Parallel_Corpus by Sanskrit Datasets.

It achieves the following results on the evaluation set:

• Loss: 0.2660
• Memory/max Mem Active(gib): 20.79
• Memory/max Mem Allocated(gib): 20.79
• Memory/device Mem Reserved(gib): 21.46

Model description

This is a fine-tuned version of Qwen2.5-VL-7B-Instruct, specifically adapted for Sanskrit OCR (Optical Character Recognition) tasks. The model has been trained using LoRA (Low-Rank Adaptation) on a dataset of Sanskrit text images and their corresponding transcriptions.

Key Features:

• Base Model: Qwen/Qwen2.5-VL-7B-Instruct (7 billion parameters)
• Task: Sanskrit OCR - converting Sanskrit text images to machine-readable text
• Training Method: LoRA fine-tuning with vision-language capabilities
• Dataset: Sanskrit OCR Parallel Corpus with chat template formatting
• Architecture: Vision-Language Model with multimodal understanding

Capabilities:

• Read and transcribe Sanskrit text from images
• Handle various Sanskrit scripts and fonts
• Process both text and visual inputs simultaneously
• Generate accurate Sanskrit text transcriptions

The model maintains the original Qwen2.5-VL's vision-language capabilities while being specialized for Sanskrit text recognition tasks.

Training and evaluation data

Training Dataset

The model was trained on the diabolic6045/sanskrit-ocr-parallel-corpus-chat-template dataset, which contains Sanskrit text images paired with their corresponding transcriptions. The dataset was converted from the original snskrt/Sanskrit_OCR_Parallel_Corpus and formatted with chat templates for vision-language training.

Evaluation Results

The model was evaluated on a test set of 314 Sanskrit text samples:

Metric	Value
Total Samples	314
Successful Samples	314
Failed Samples	0
Success Rate	100.00%
Exact Match Accuracy	1.59%
Character-level Accuracy	86.38%
Token-level Jaccard Similarity	50.44%

Key Insights:

• The model successfully processes all test samples without failures
• High character-level accuracy (86.38%) indicates good recognition of individual Sanskrit characters
• Lower exact match accuracy (1.59%) suggests room for improvement in complete text transcription
• Moderate token-level similarity (50.44%) shows reasonable semantic understanding

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0002
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• gradient_accumulation_steps: 4
• total_train_batch_size: 16
• total_eval_batch_size: 4
• optimizer: Use OptimizerNames.ADAMW_BNB with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 110
• training_steps: 1105

Training results

Training Loss	Epoch	Step	Validation Loss	Mem Active(gib)	Mem Allocated(gib)	Mem Reserved(gib)
No log	0	0	3.3372	17.59	17.59	17.66
0.2428	1.0	369	0.3075	20.69	20.69	21.27
0.2057	2.0	738	0.2660	20.79	20.79	21.46

This model was trained using:

See axolotl config

axolotl version: 0.12.2

base_model: Qwen/Qwen2.5-VL-7B-Instruct
processor_type: AutoProcessor

# these 3 lines are needed for now to handle vision chat templates w images
skip_prepare_dataset: true
remove_unused_columns: false
sample_packing: false

chat_template: qwen2_vl
datasets:
  - path: sanskrit_multimodal_train.json
    type: chat_template
    field_messages: messages
dataset_prepared_path: last_run_prepared
val_set_size: 0.01
output_dir: ./outputs/out-qwen2-5-vl

adapter: lora
lora_model_dir:

sequence_len: 2048
pad_to_sequence_len: false

lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_target_modules: 'model.language_model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'

wandb_project: Sanskrit-OCR
wandb_entity:
wandb_watch:
wandb_name: qwen2-5-vl-sanskrit-ocr
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 2
num_epochs: 3
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0002

bf16: true
fp16:
tf32: true

gradient_checkpointing: true
logging_steps: 1
flash_attention: true
eager_attention:

warmup_ratio: 0.1
evals_per_epoch: 1
saves_per_epoch: 1
weight_decay: 0.0

# Automatically upload checkpoint and final model to HF
hub_model_id: diabolic6045/qwen2-5-vl-sanskrit-ocr-lora
# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

Framework versions

• PEFT 0.17.0
• Transformers 4.55.2
• Pytorch 2.7.1+cu128
• Datasets 4.0.0
• Tokenizers 0.21.2