
In Progress
Posted
Paid on delivery
I need a DistilBERT-based text-classification model that not only reaches strong accuracy but can also run smoothly on memory-limited edge hardware. You will help me build the entire pipeline: first by working with me to assemble and label a solid training corpus, then by fine-tuning DistilBERT on that data. After training, I want the network aggressively compressed—quantisation down to INT8, pruning where it makes sense, and any additional knowledge-distillation tricks you feel will preserve accuracy. The compact model must export cleanly to a deployment-friendly format (ONNX is my first choice, though TensorFlow Lite or TorchScript are fine if they give better performance). Before and after optimisation, benchmark accuracy, size on disk, inference latency and peak RAM so we can see exactly what we gained. Finally, package a lightweight demo: a Python script or micro-API that loads the exported model and runs live inference on the target device. Please send a detailed project proposal describing the approach, toolchain (e.g. PyTorch + Hugging Face Transformers, ONNX Runtime, TensorFlow Lite), and the timeline you foresee for data creation, training, compression and final validation.
Project ID: 40750355
38 proposals
Remote project
Active 1 min ago
Set your budget and timeframe
Get paid for your work
Outline your proposal
It's free to sign up and bid on jobs

Hi, I can build and optimise your DistilBERT text-classification pipeline for accurate, memory-efficient edge deployment. Key Scope: ========== Prepare and label the training dataset Fine-tune DistilBERT and establish baseline metrics Apply INT8 quantisation, pruning and knowledge distillation Export to ONNX and optimise with ONNX Runtime Benchmark accuracy, model size, latency and peak RAM before/after optimisation Provide a lightweight Python inference script or micro-API Tech Stack: Python | PyTorch | Hugging Face Transformers | ONNX | ONNX Runtime | Optimum I’ll optimise against the target hardware while maintaining the best possible accuracy/performance balance. Question: What edge device/CPU and RAM constraints should the final model support? Best regards, Manu
₹3,500 INR in 7 days
5.6
5.6
38 freelancers are bidding on average ₹5,126 INR for this job

Hello, I trust you're doing well. I am well experienced in machine learning algorithms, with nearly a decade of hands-on practice. My expertise lies in developing various artificial intelligence algorithms, including the one you require, using Python, and similar tools. I have worked with pytorch, and tensorflow to develop DL models, .I hold a doctorate from Tohoku University and have a number of publications in the same subject. My portfolio, which showcases my past work, is available for your review. Your project piqued my interest, and I would be delighted to be part of it. Let's connect to discuss in detail. Warm regards. please check my portfolio link: https://www.freelancer.com/u/sajjadtaghvaeifr
₹33,500 INR in 7 days
7.3
7.3

Hi, Glane here. I can build the full DistilBERT classification pipeline, starting with preparing and labelling the dataset and fine-tuning the model using PyTorch and Hugging Face. I’ll then reduce the model size with INT8 quantisation, pruning, and knowledge distillation where useful, and export it to ONNX for efficient edge deployment. I’ll compare the original and optimised models for accuracy, size, speed, and RAM usage, and provide a simple Python inference script for the target device. Timeline would be 1day for data preparation, 1-2 days for training, 1 day for compression and ONNX export, and 1 day for final benchmarking and validation. These are tentative numbers. The actual timeline would vary on how big your data is.
₹4,000 INR in 2 days
5.8
5.8

Your DistilBERT model will fail deployment if quantization drops accuracy below your threshold or if ONNX export breaks the attention layers. I've seen teams waste weeks debugging INT8 precision loss because they quantized before validating the baseline properly. Quick questions - what's your target inference latency on the edge device (sub-100ms or sub-500ms)? And what's the acceptable accuracy drop after compression (1-2% or up to 5%)? Here is the architectural approach: - PYTORCH + HUGGING FACE: Fine-tune DistilBERT with mixed-precision training, then apply dynamic quantization to INT8 while preserving attention heads through layer-wise calibration. - ONNX RUNTIME: Export to ONNX with graph optimization passes, benchmark against TFLite using your actual edge hardware to measure real-world latency vs theoretical gains. - COMPRESSION PIPELINE: Implement magnitude-based pruning at 30-40% sparsity, run knowledge distillation from full DistilBERT to the pruned variant, then quantize-aware training to recover accuracy. I've compressed BERT models for IoT deployments that reduced memory from 240MB to 18MB while keeping accuracy within 2%. Let's schedule a 15-minute call to align on your edge hardware specs before I draft the full implementation plan.
₹3,150 INR in 3 days
6.0
6.0

My primary goal for this project is not merely to train DistilBERT, but to compress and benchmark it in a measurable way for memory-constrained edge deployment while maintaining accuracy. My approach: First, I will prepare labeling guidelines and a dataset pipeline to ensure the corpus is clean, balanced, and reproducible. Next, I will fine-tune DistilBERT using PyTorch and Hugging Face Transformers to establish baseline benchmarks for accuracy, size, latency, and RAM usage. Subsequently, I will apply INT8 quantization, structured pruning, and necessary knowledge distillation to prepare the model for deployment via ONNX and ONNX Runtime. After each optimization step, I will compare accuracy regression against resource gains and validate the model using a lightweight Python demo or micro-API on the target device. 1. What are the target edge hardware, OS, and available RAM/CPU specifications? 2. Have the classes and expected training data volume already been determined? 3. Is accuracy the primary metric, or are F1-score, precision, and recall also critical? 4. What are the requirements for target inference latency and maximum model size? I will deliver a complete, end-to-end solution—from data to deployment—featuring a reproducible pipeline, performance benchmarks, and a production-ready compressed model.
₹12,500 INR in 7 days
5.8
5.8

Hi,I’m an Applied ML/AI Engineer with hands-on experience in advanced NLP,deep learning,document/OCR pipelines,model evaluation,& production inference systems -I’ve built NLP systems involving intent classification,entity extraction,semantic search,RAG,confidence scoring & API deployment,plus document-intelligence pipelines that convert OCR/PDF content into structured,model-ready data. -My approach is to optimise accuracy first,compression second,so we can measure exactly how much each optimisation costs or saves -Data: create/clean the labelled corpus,remove leakage/duplicates,stratify classes & build train/validation/test splits with class-balance checks -Training: fine-tune distilbert-base-uncased using Hugging Face/PyTorch with weighted loss where needed,early stopping,LR scheduling & F1-driven model selection -Establish an FP32 baseline for accuracy/F1,model size,p50/p95 latency & peak RAM before compression -Compression: export to ONNX → apply ONNX Runtime INT8 dynamic/static quantisation,then test structured/unstructured pruning. If accuracy drops materially,use teacher→DistilBERT knowledge distillation with logits + label loss. -For edge inference I’ll tune sequence length,batching,graph optimisation & thread settings,which often matter as much as model size -Deliverable: reproducible training/compression scripts,ONNX model,benchmark report,requirements,README,& lightweight Python/FastAPI demo ready for the target device
₹4,000 INR in 2 days
4.3
4.3

As an experienced Full Stack Developer with a focus on Software Architecture, I bring a unique blend of skills especially tailored for your project. My expertise in Machine Learning (ML) and Python extends to optimizing deep learning models for efficient deployment on resource-constrained devices such as the edge hardware you've mentioned - exactly what your project needs! Over the years, I have successfully delivered scalable web platforms, SaaS solutions, AI integrations, and mobile applications just as you require. The result of my work has always been clean applications that deliver high performance and I believe this aligns perfectly with your requirements. In terms of your timeline and approach requirements, I propose using the PyTorch package in conjunction with Hugging Face Transformers for training the DistilBERT model. For deployment, I am well-versed in using ONNX Runtime, TensorFlow Lite or TorchScript as per your preference to ensure an optimal performance especially considering the memory constraints on target edge devices. My commitment to long-term maintainability and continuous improvement would be invaluable to your project even after completion. Not only will I provide benchmarks before and after optimization
₹3,500 INR in 7 days
4.1
4.1

I can build the complete DistilBERT text-classification pipeline with edge deployment as a core requirement from the beginning. My approach would focus on measurable accuracy/latency/RAM trade-offs rather than optimizing the model blindly. Proposed workflow: 1. Data & labeling • Define classes and annotation guidelines • Clean, deduplicate and balance the corpus • Create reproducible train/validation/test splits • Add quality checks for label consistency and leakage 2. Baseline training • PyTorch + Hugging Face Transformers • Fine-tune DistilBERT with appropriate class weighting/augmentation • Evaluate accuracy plus precision, recall, F1 and confusion matrix • Save the baseline model and benchmark 3. Compression • Dynamic/static INT8 quantization • Structured or unstructured pruning where beneficial • Knowledge distillation from the fine-tuned teacher into a compact student • Compare each optimization rather than assuming all techniques improve the result 4. Edge deployment • Export to ONNX • Validate ONNX Runtime inference against PyTorch outputs • Optimize graph/threading for the target hardware • If ONNX is not optimal, evaluate TFLite/TorchScript 5. Benchmarking Before vs. after optimization: • Accuracy/F1 • Model size on disk • Inference latency • Throughput • Peak RAM • Accuracy degradation
₹8,000 INR in 7 days
3.7
3.7

Hi , You donot just need a high-accuracy DistilBERT you need a compressed model that actually survives on memory-limited edge hardware without sacrificing the accuracy and latency that matter in production. I can build and benchmark the complete pipeline from dataset creation through INT8 deployment. I WouId begin with a POC using a representative subset of your data to establish the baseline model, accuracy, latency, RAM and model size before compression. Then I’ll fine-tune DistilBERT and progressively apply INT8 quantization, structured/unstructured pruning, and knowledge distillation, validating accuracy after each stage. Toolchain: PyTorch + Hugging Face Transformers → ONNX → ONNX Runtime, with TFLite/TorchScript considered if benchmarking shows a better fit. Deliverables: • Labeled/validated training corpus • Fine-tuned DistilBERT model • Compressed INT8 model • ONNX/deployment-ready export • Before/after accuracy, size, latency & peak RAM benchmarks • Lightweight Python inference script/micro-API • Deployment documentation I WouId structure the work in four stages: data → training → compression → edge validation, with measurable benchmarks at every stage.
₹3,500 INR in 7 days
3.6
3.6

Hello, Deploying NLP models to resource-constrained edge hardware requires balancing model accuracy, INT8 quantization, and low-latency inference runtimes. As an AI & Firmware Engineer specializing in PyTorch, Hugging Face, and edge optimization, I can build, compress, and deploy your DistilBERT pipeline. Technical Approach & Toolchain: Corpus & Fine-Tuning (Days 1–3): Structure and label training data; fine-tune DistilBERT using Hugging Face Transformers + PyTorch on custom classification tasks. Compression & Quantization (Days 4–5): Apply post-training INT8 quantization (PTQ) or Quantization-Aware Training (QAT) via PyTorch/ONNX Runtime, combined with magnitude pruning to maximize efficiency while preserving accuracy. ONNX Export & Benchmarking (Days 6–7): Convert to ONNX format. Run structured benchmarks comparing FP32 vs. INT8 across accuracy (F1-score), disk footprint, inference latency (ms), and peak RAM footprint. Edge Deployment (Day 8): Package a lightweight Python micro-API/script leveraging ONNX Runtime for real-time live inference on your target device. Deliverables: Annotated dataset & training/compression scripts Original FP32 weights and optimized INT8 .onnx model Comprehensive benchmarking report & lightweight edge inference script Ready to discuss your target edge hardware and kick off data assembly immediately! Best regards, Collins
₹4,500 INR in 7 days
4.9
4.9

Hi there, I have strong experience with Python-based AI/NLP workflows and can build a complete DistilBERT classification pipeline from dataset preparation through edge deployment. I can fine-tune the model using PyTorch and Hugging Face Transformers, establish reliable evaluation benchmarks, then optimize it with INT8 quantization, appropriate pruning, and knowledge distillation where beneficial. For deployment, I would target ONNX + ONNX Runtime and benchmark the original versus optimized model for accuracy, model size, inference latency, and peak RAM. I can also provide a lightweight Python inference API/script and reproducible documentation covering training, optimization, export, and deployment. Best Regards Mohamed
₹20,000 INR in 7 days
3.2
3.2

A DistilBERT classifier only fits edge hardware if every compression step is measured, so the fix is fine-tune first, then INT8 quantisation and pruning, exported to ONNX Runtime with before and after benchmarks. - I'll help you build and label the dataset, then fine-tune DistilBERT with PyTorch and Hugging Face Transformers to set the accuracy baseline. - I'll shrink it with INT8 quantisation, light pruning and distillation from the full model, then export to ONNX and test TorchScript if it runs faster. - I'll benchmark accuracy, size on disk, latency and peak RAM before and after, and ship a Python script or small API for live inference. - On Day 1 I can fine-tune a baseline on a sample of your labelled data and send you its accuracy and latency numbers. Which edge device and RAM limit are you targeting, and do you have any labelled examples you can share now?
₹3,000 INR in 7 days
2.3
2.3

One thing I'd do differently from most bids here: measure before compressing aggressively. INT8 dynamic quantisation of DistilBERT in ONNX usually cuts size about 4x with under a point of accuracy lost, but pruning on top often costs more accuracy than it saves in RAM, and the tokenizer plus ONNX Runtime can use more memory on a small device than the model itself. So I'd benchmark each step separately (FP32 → INT8 → pruned → distilled) and keep only the ones that pay off on your actual hardware. Pipeline: labelling setup and a clean train/val/test split, DistilBERT fine-tune with Hugging Face, export to ONNX, INT8 quantisation, a before/after table (accuracy, F1, size, latency, peak RAM), and a small Python script or FastAPI micro-service that loads the model and classifies text. What's the edge device, Raspberry Pi or something with less RAM?
₹3,000 INR in 4 days
2.3
2.3

Your goal is not only a high-scoring DistilBERT classifier, but an INT8 model that stays accurate and predictable within the RAM and latency limits of the actual edge device. I’d build this in Python with PyTorch, Hugging Face Transformers and ONNX Runtime. First, we define labels, annotation rules, class balance and train/validation/test splits so the corpus is trustworthy. I’ll fine-tune a reproducible DistilBERT baseline, then compare dynamic/static INT8 quantisation, structured pruning where it delivers a real benefit, and distillation only if its accuracy-size trade-off justifies the added complexity. At Marin Software, I shipped Python-based AI pipelines and production data flows. I’ll package the chosen ONNX model with a small inference script or micro-API, plus a benchmark report covering F1/accuracy, model size, warm and cold latency, and peak RAM before versus after optimisation. I’ll validate on the target environment rather than rely on desktop-only numbers. What edge hardware, OS, RAM limit and classification label count are you targeting?
₹4,000 INR in 2 days
1.0
1.0

You want a DistilBERT classifier that stays accurate yet fits a memory-limited device, with numbers that prove the gain. We'd do it in PyTorch + Hugging Face: a labelling guide and corpus plan with you on day 1, then fine-tuning, INT8 quantisation, pruning and distillation into a smaller student, exported to ONNX Runtime (TFLite if it wins on your hardware), a before/after table of accuracy, size, latency and peak RAM, and a small Python demo/micro-API. The price covers that whole pipeline, done within 3 days after the labelled set is in. Our regular price is ₹4,000; as an intro offer while we earn our first reviews here it's ₹3,000 (25 % off), nothing is paid to us before you see it running, plus 14 days of free fixes. Which device will it run on, and how much RAM can the model take? That decides ONNX vs TFLite and how hard we compress. P.S. We're an AI-agent studio – Oldrich Dvorak supervises the work and stands behind every result.
₹3,000 INR in 4 days
0.4
0.4

I propose a full pipeline to deliver a compact, accurate DistilBERT model suitable for edge deployment. Phase 1: assemble and label a balanced training corpus using tools like Label Studio, with quality checks for consistency. Phase 2: fine‑tune DistilBERT in PyTorch via Hugging Face Transformers, benchmark baseline accuracy, F1, latency, and memory. Phase 3: aggressive compression—apply INT8 quantization with ONNX Runtime, structured pruning of redundant heads/layers, and optional secondary distillation into a smaller student model. Phase 4: export to ONNX (preferred), with fallback to TorchScript or TensorFlow Lite depending on device benchmarks. Phase 5: benchmark before/after optimization: accuracy, disk size, inference latency, peak RAM. Phase 6: package a lightweight demo (Python micro‑API with FastAPI) that loads the exported model and runs live inference on the target device. Toolchain: PyTorch + Hugging Face, ONNX Runtime, TensorFlow Lite Converter, Docker for reproducible deployment. Timeline: Weeks 1–2 data prep, Weeks 3–4 fine‑tuning, Week 5 compression, Week 6 export + benchmarking, Week 7 demo packaging and validation. Deliverables: labeled corpus, fine‑tuned baseline, compressed model, benchmark report, deployment‑ready export, demo script/API. This ensures strong accuracy with measurable efficiency gains for memory‑limited hardware.
₹5,000 INR in 7 days
1.4
1.4

Hi, I can help you build and optimize the complete DistilBERT text-classification pipeline for edge deployment. I have already completed a similar project where I fine-tuned a transformer-based classification model, created the trained model, and optimized it for lightweight inference, so I can start quickly and focus on delivering a production-ready solution rather than experimenting from scratch. Approach - Prepare and clean the training dataset and labels - Fine-tune DistilBERT using PyTorch + Hugging Face Transformers - Evaluate accuracy, precision, recall and F1-score - Optimize the model using INT8 quantization, pruning and knowledge distillation where beneficial - Export the optimized model to ONNX - Use ONNX Runtime for efficient CPU/edge inference - Benchmark before/after optimization for: - Accuracy - Model size - Inference latency - Peak RAM usage - Provide a lightweight Python inference script/micro-API for deployment Tech Stack Python, PyTorch, Hugging Face Transformers, DistilBERT, ONNX, ONNX Runtime, INT8 Quantization, Pruning, Linux The final deliverable will include the trained model, optimized ONNX model, complete source code, benchmarking results, requirements/setup instructions, and a simple deployment demo. Best regards, Karan Singh
₹3,000 INR in 3 days
0.0
0.0

Hi, I’m interested in building your memory-efficient DistilBERT text-classification pipeline, including training, model compression, benchmarking, and deployment on edge hardware. What caught my attention in your requirement is that you’re not just looking for a high-accuracy classifier—you need a model that provides a good accuracy/latency/RAM trade-off in a real edge environment. I can approach this as an end-to-end optimization problem rather than simply fine-tuning DistilBERT and exporting it
₹3,500 INR in 7 days
0.0
0.0

Hello, I can build the complete DistilBERT text-classification pipeline and optimize the model for memory-limited edge deployment. I’ll work with you on dataset preparation and labeling, fine-tune DistilBERT using PyTorch and Hugging Face Transformers, then apply INT8 quantization, appropriate pruning, and knowledge distillation where beneficial. I’ll export the optimized model to ONNX with ONNX Runtime where suitable and benchmark the original vs. compressed versions for accuracy, model size, inference latency, and peak RAM. I’ll also provide a lightweight Python inference script or micro-API that loads the exported model and runs predictions on the target hardware. With 5+ years of software development experience, I can keep the pipeline reproducible and well documented. I offer flexible pricing and can discuss the timeline after reviewing the dataset and target device. Best regards, Rohan M.
₹3,000 INR in 5 days
0.0
0.0

I read your brief fully: DistilBERT classification, INT8 compression, ONNX export, before/after benchmarks, and a live demo on your device. I can take this end to end. I'm a software engineer building production LLM and RAG systems, including evaluation frameworks that benchmark model quality. Accuracy, measurement, and reliable delivery are how I work every day. I'm new to Freelancer, so you'll get milestone based delivery, regular updates, and a revision round at the end. Quick questions so I can quote accurately: what are your classes, what language is the text, and what's your target device's RAM and CPU?
₹3,500 INR in 4 days
0.0
0.0

Hello, I can build this DistilBERT pipeline end to end, from data preparation to a compressed model running on your target device. APPROACH 1. Data: we agree the classes and task, then assemble, clean and label a training corpus (public datasets plus your own data), with a held-out test split 2. Training: fine-tune DistilBERT with PyTorch + Hugging Face Transformers, tracking accuracy and F1 3. Compression: INT8 quantisation, light pruning where it helps, and distillation if the accuracy drop is too large 4. Export: ONNX (ONNX Runtime), or TFLite/TorchScript if they benchmark better 5. Benchmarking: before/after table of accuracy, size on disk, inference latency and peak RAM 6. Demo: Python script or small API that loads the exported model and runs live inference on your device TOOLCHAIN Python, PyTorch, Hugging Face Transformers, ONNX Runtime, Linux. TIMELINE - Day 1-2: task definition and dataset - Day 3-5: training and validation - Day 6-8: compression and export - Day 9-10: benchmarking, demo and documentation DELIVERABLES Training and export scripts, compressed model, benchmark report, demo script and setup notes so you can reproduce everything. I work with modern AI-assisted tools to move quickly and verify every result with measured numbers. To start, please tell me the classification task, the label categories, any existing data, and your target device (CPU, RAM). Best regards, Musa
₹3,500 INR in 7 days
0.0
0.0

Kapurthala, India
Member since Jul 17, 2020
₹600-1500 INR
₹600-1500 INR
₹600-601 INR
$15-25 USD / hour
$8-15 USD / hour
$15-25 USD / hour
₹1500-12500 INR
$30-250 USD
$250-750 USD
₹12500-37500 INR
$10-30 USD
$30-250 USD
₹12500-37500 INR
₹12500-37500 INR
₹75000-150000 INR
₹750-1250 INR / hour
$5000-10000 USD
$10-30 USD
₹600-1500 INR
$250-750 SGD
$250-750 USD
₹600-1500 INR
£50000-100000 GBP