You can deploy this model and start earning money today!

akash1341/
lr5.0e-05_sched-constant_with_warmup_acc5_ebs5_warm0.05

Name: lr5.0e-05_sched-constant_with_warmup_acc5_ebs5_warm0.05
Availability: InStock

peft

safetensors

Views

No views yet

OLMOCR Fine-Tuned Model (LoRA Adapter)

This repository contains the fine-tuned PEFT/LoRA adapter weights for the OLMOCR model, targeted specifically at advanced optical character recognition tasks.

Model Details

Model Description

This model acts as a custom LoRA adapter designed to work seamlessly with the allenai/olmocr-7b-instruct base model. It is optimized to transcribe images with exact structural accuracy using standard prompts without extra elaborative text.

Developed by: Bathini Akash / David Smith
Model type: Causal Language Model (LoRA Fine-tuning)
Base model: allenai/olmocr-7b-instruct
Finetuning approach: PEFT / LoRA (rank: 16, alpha: 32)
Target modules: up_proj, q_proj, o_proj, down_proj, gate_proj, v_proj, k_proj

Uses

Direct Use

To use this adapter, you will need a base instance of allenai/olmocr-7b-instruct. The model is generally served via vLLM using the multi-modal Lora serving capabilities.

For inference, a standard system prompt should request:

"Extract the exact text from the image. No explanation."

Environmental Details & Inference

This model is highly optimized to run with vLLM dynamically. Hardware testing typically utilizes A100 GPU nodes with memory utilization limits (e.g. 0.60) in float16 formats.

Training Details

Training Hyperparameters

Learning Rate: 5e-05
Scheduler: constant_with_warmup
Warmup Ratio: 0.05
Gradient Accumulation Steps: 5
Epochs: 3
Dataset length: 92,110 examples

Training Framework

PEFT 0.13.2
vLLM
PyTorch

Evaluation

The model's results are validated via a dynamic evaluation script where the outputs are strictly compared side-by-side with ground-truth representations using Levenshtein distance matching and qualitative failure checking on dynamic context lengths up to 15,000 tokens on test dataset subsets.

akash1341/lr5.0e-05_sched-constant_with_warmup_acc5_ebs5_warm0.05