Use cases
Where a small model beats a big one
Tasks that are narrow, repeated many times a day, and already have examples of good answers. Each one has a template in the builder that sets the prompt format, training defaults and how the switch is checked.
Classification and routing
One label per input, at the volume your queue actually has.
Support ticket routing · Intent detection for assistants and IVR · Triage and priority
Read moreStructured extraction
Documents in, JSON that matches your schema out.
Invoice and receipt fields · Insurance claims · Contract clauses and dates
Read moreTagging
Every tag that applies, zero or more per input.
Topic and theme tags · Compliance and risk flags · Product attributes
Read moreRedaction
Remove sensitive details, keep everything else exactly as it was.
Call and chat transcripts · Clinical and case notes · Support tickets
Read moreSummarization
Short, faithful summaries in the format your team reads.
Call and meeting notes · Support case summaries · Ticket and thread digests
Read moreRegulated industries
Training data, compute and weights stay in your AWS account.
Compute runs on EC2 instances in your account and terminates when the run ends · Datasets, checkpoints and model weights are written only to your S3 bucket · Access is one IAM role with a per-organization External ID, revocable by deleting it
Read moreWhy teams switch
General models are a costly way to do one narrow job
Most LLM traffic in production is the same task repeated: route this, extract that, redact the rest. A small model trained on your own examples does that one job, on hardware you control.
Per-token bills grow with every request
A routing label or a JSON record is a few tokens of output, but every call pays to send the whole prompt to a general-purpose model. A GPU you run costs the same per hour however many requests it serves.
Latency and limits you don't control
Hosted models add network round trips, rate limits and model deprecations to your critical path. A small model on your own instance answers in milliseconds and changes only when you retrain it.
Your records leave your environment
Every API call sends customer data to another company. Training and serving in your own AWS account keeps it where your security team already reviewed it.
The hard parts are handled
Fine-tuning is mostly infrastructure and data work. SynqTune does it for you, in your account.
GPU provisioning
Picks the instance, requests spot capacity, falls back to on-demand, and terminates it afterwards.
Data cleanup
Deduplication, filtering, PII scrubbing and chat formatting as configurable steps.
Checkpointing
Saves progress to S3 so a reclaimed spot instance does not cost you the run.
Evaluation
Scores the model on data it never trained on, and against the LLM you use today.
Packaging
Writes the adapter, a model card with checksums and an optional GGUF file.
Run visibility
Step-by-step progress, training loss and logs while the run is in progress.
Data preparation
Clean training data without writing scripts
Most of a model's quality comes from its data. Each preparation step is a switch with sensible defaults, and the result is a Parquet file in your bucket that you can inspect.
- Remove exact and near-duplicate records with MinHash
- Filter by language, length and repetition; drop refusals and AI disclaimers
- Redact emails, phone numbers, SSNs, card numbers and names before training
- Optionally grade responses with Claude, OpenAI or Gemini using your own key and drop weak ones
- DeduplicateExact hash or MinHash near-duplicates
- Quality & language filterLanguage, length, repetition
- PII scrubEmails, phone numbers, SSNs, cards, names
- Response quality checkEmpty, refusals, disclaimers, invalid JSON
- Grade with an LLMOptional · Claude, OpenAI or Gemini
- Format conversionChatML, Alpaca or ShareGPT
Training
LoRA fine-tuning on spot GPUs, without babysitting
SynqTune launches a GPU instance in your account, trains a LoRA adapter with Unsloth, and shuts the instance down when it is done.
- Spot capacity by default with automatic on-demand fallback
- Checkpoints saved to your S3 so interrupted runs resume, not restart
- Live step progress, training loss and logs on the run page
- Instances terminate when a run finishes, fails or stops reporting
Checkpoint
Saved to your S3 during training
Interrupted
Spot capacity reclaimed by AWS
Resumed
Continues from the last checkpoint
Comparison report
Know it's good enough before you switch
The fine-tuned model and the LLM you use today answer the same held-out examples from your data. The report scores both for your task and ends with a verdict.
- Scored for the task: accuracy for labels, field accuracy and schema checks for JSON, leak rate for redaction, a blind judge for summaries
- A breakdown per label, field or tag, so a weak spot is visible before it reaches production
- Latency and cost per 1,000 requests for both models, from measured runs
- Ready to replace, close, or not ready, with the reasons and the examples behind them
Export
Standard files you can run anywhere
No proprietary format and no lock-in. Everything lands in your bucket in a predictable layout, ready for your serving stack.
- Hugging Face PEFT adapter in safetensors, loadable with transformers
- 4-bit Q4_K_M GGUF for llama.cpp and Ollama, one click to download
- Model card and README with settings, evaluation and SHA-256 checksums
- slm-customer-bucket-<account-id>/
- raw/— your uploads
- data/— prepared Parquet per run
- model/<run>/
- adapter_model.safetensors— LoRA adapter
- adapter_config.json
- model_card.json— settings + eval + SHA-256
- README.md— how to load it
- gguf/<name>.Q4_K_M.gguf— llama.cpp · Ollama
Pipeline steps
Every step available today
Combine these in the pipeline builder. Each one runs on an instance in your account and writes its output to your bucket.
Sources
Amazon S3
Read JSONL, JSON, CSV, TSV or Parquet from any path the role can access
Direct upload
Browser uploads straight into your bucket with a presigned URL
Data preparation
Deduplication
Exact or MinHash near-duplicate removal
Quality & language filter
Language detection, length and repetition limits
Tabular + PII scrub
Maps columns to chat turns and redacts personal data
Response quality check
Drops refusals, disclaimers, echoes and invalid JSON
Data augmentation
Adds varied copies of short datasets
Format converter
ChatML, Alpaca or ShareGPT
Tokenizer training
Trains a tokenizer on your corpus
LLM grading (optional)
Grade with Claude
Drops responses that miss your criteria
Grade with OpenAI
Same step, your OpenAI key
Grade with Gemini
Same step, your Google key
Train, evaluate, ship
LoRA fine-tuning
Unsloth on a GPU instance, with S3 checkpoints
Evaluation
Held-out scores, plus an optional comparison with your current LLM
Save model
Adapter, model card, SHA-256 checksums, optional GGUF
FAQ
More questions
More on setup, every pipeline step and troubleshooting is in the documentation.
Which models can I fine-tune?
Open models supported by Unsloth, typically from 1B to 8B parameters, such as Gemma, Llama, Mistral, Phi and Qwen. Templates pick a small default; you can enter any Hugging Face model ID under Advanced.
What happens if a spot instance is interrupted?
Checkpoints are written to your bucket during training. When the run is resumed it continues from the last checkpoint instead of starting over.
How do I remove SynqTune's access?
Delete the CloudFormation stack, or just the IAM role. SynqTune can no longer act in your account, and the S3 bucket and its contents are kept.