Parsyn Documentation

Fine-tune language models on your own terms. Bring your GPUs or use ours. Keep full control over your data and training pipeline.

What is Parsyn

Parsyn is a platform for fine-tuning large language models. You upload a dataset, pick a base model, configure your training parameters, and Parsyn handles orchestration, real-time monitoring, checkpointing, and inference.

What makes Parsyn different is the compute model. You have two options:

• Use Parsyn's GPU fleet: No hardware to manage. Pick a GPU type, pay with credits, and start training. The platform manages the infrastructure.
• Bring your own GPUs: Install the Parsyn worker on your machines. They connect to the platform and receive tasks automatically. You keep full control over your hardware.

Both options work through the same interface. You can even mix them: train on your own GPUs during business hours and overflow to Parsyn's fleet when you need extra capacity.

Who is it for

• Data scientists and ML engineers who want to fine-tune open models (Llama, Mistral, Phi, etc.) on proprietary data without writing training infrastructure from scratch.
• Teams with existing GPU hardware who want a centralized platform to manage training jobs, datasets, and models across their machines.
• Companies without GPUs who want to fine-tune models without setting up cloud GPU instances manually. Credits-based pricing, no commitment.
• Research teams that need reproducible training runs with full parameter tracking, checkpoint management, and evaluation tools.

How it works

The platform follows a hub-and-spoke architecture. The Parsyn platform is the hub: it stores all metadata, coordinates workers, and serves the dashboard. Workers are the spokes: they handle the actual GPU compute.

┌─────────────────────────────────────────────────────┐
│           Parsyn Platform (parsyn.progatis.com)       │
│                                                       │
│   Dashboard, API, job orchestration, billing,        │
│   real-time monitoring, dataset/model storage        │
│   PostgreSQL + Redis + S3                            │
└───────────────────────┬───────────────────────────────┘
                        │ WebSocket (bidirectional)
          ┌─────────────┼─────────────┐
          │             │             │
   ┌──────▼──────┐┌────▼──────┐┌─────▼────────┐
   │ Parsyn GPU  ││ Your GPU  ││ Your GPU     │
   │ (managed)   ││ Server A  ││ Server B     │
   │ Credits/hr  ││ (free)    ││ (free)       │
   └─────────────┘└───────────┘└──────────────┘

Workers are stateless. If a worker disconnects (network issue, machine reboot), the platform detects it through missed heartbeats. Training checkpoints are saved to S3, so no progress is lost.

Parsyn Workers vs. Your Workers

Key features

Training

• Standard fine-tuning and LoRA/QLoRA adapter training
• Federated learning across multiple workers
• Automatic checkpointing with S3 storage
• Mixed precision (fp16/bf16) and gradient accumulation
• Real-time metrics streaming to the dashboard (loss, learning rate, GPU utilization, ETA)
• Smart worker assignment based on GPU benchmarks
• Training presets for quick configuration (pre-built hyperparameter templates)
• Training pipelines: automated multi-step workflows (train → evaluate → export)

Dataset management

• JSONL, CSV, and Parquet formats
• Multipart upload for files up to 10 GB
• Automatic validation, statistics, and schema detection
• Built-in split, deduplication, format conversion, and normalization
• Synthetic data generation: generate, augment, paraphrase, or filter datasets using LLMs

Model management

• Import from HuggingFace Hub by model ID
• Upload custom models (safetensors, PyTorch checkpoints)
• Version tracking across fine-tuning runs
• Post-training pipeline: LoRA merge, quantization (4-bit, 8-bit)
• Export to inference formats: GGUF (llama.cpp), ONNX, GPTQ, AWQ quantization

Inference and evaluation

• Built-in chat interface with streaming responses
• A/B model comparison
• Arena mode with ELO ratings
• Automated evaluation (perplexity, custom metrics)
• Evaluation suites: structured evaluation with predefined prompt sets, scoring, and cross-model comparison

Operations

• Subscription plans with quotas + credit-based billing (Stripe)
• Organizations and teams with project-level resource isolation
• Multi-channel notifications (email, Slack, Telegram, Teams)
• Worker fleet management with master enrollment keys
• Role-based admin panel
• Two-factor authentication (TOTP, email, backup codes)

Tech stack

Next steps

• Quickstart to run your first training job in under 10 minutes
• Core Concepts to understand how datasets, models, jobs, and workers fit together
• Worker Setup to connect your own GPU machines to the platform