- What Is Fine-Tuning? How It Works and When You Need It. How it differs from prompting and RAG, and when it earns its cost.
- Air-Gapped LLM Training: What It Actually Requires. What a real air gap demands, and what "private cloud" usually means instead.
- Synthetic Data for Fine-Tuning When You Have No Labels. Getting a training set when nobody labelled anything.
- SFT vs. RL for Agents: Reinforcement Learning Fine-Tuning. Which technique fits which failure mode.
- Financial Text Classifiers: Fine-Tuning vs. Prompting a Small LLM. Benchmarks on seven tasks: LoRA vs. prompting, TF-IDF, and full tuning.
- Fine-Tuning LLMs for Finance: A Practical Guide With Open-Source Tools. A worked example on financial data.
- Training a Language Model for Internal Wikis. Training against the wikis and runbooks your team already wrote.
- Choosing a base model to train from: guides to the Llama, Qwen, Mistral, DeepSeek, Gemma, and Kimi families.
AI experiments.
Under control.
P95 is Numerata's agentic training platform: manage every training job from anywhere, from a single signal model to hundreds of backtests running in parallel.
PART OF THE NUMERATA STACK · p95.run ↗
GET STARTED
Ships as an agent skill. Add it and it instruments your experiments automatically.
AGENT SKILL
Your coding agent installs it and wires up instrumentation on its own, no manual setup, no config files to write by hand.
PYTHON SDK
Prefer to wire it up yourself? One import into your existing code, and it works with PyTorch, JAX, TensorFlow, or plain Python.
BUILT FOR THE FULL WORKFLOW
From a single run to hundreds of parallel jobs, P95 scales with you.
HYPERPARAMETER SWEEPS
Define a search space and let P95 schedule trials automatically. Grid, random, or Bayesian search. Ideal for tuning signal and risk models across market regimes as results stream in.
CLOUD JOBS
Submit training jobs to the cloud directly from your CLI or SDK. Monitor progress, stream logs, and compare results against previous runs, all in one place.
TEAM COLLABORATION
Share projects and runs across your desk or research team, and compare experiments side by side.
INSTRUMENT ONCE, SCALE ANYWHERE
Wrap your training loop once. The same code runs locally, in a sweep, or as a cloud job.
UNIFIED TRACKING
Configs, metrics, and tags all land in one place, no separate tools to reconcile.
AUTOMATIC GROUPING
Runs are grouped under their sweep automatically, no manual tagging required.
FRAMEWORK AGNOSTIC
Works with PyTorch, JAX, TensorFlow, or plain Python, no rewrite required.
WHY FINE-TUNE
In head-to-head evaluations, fine-tuned classifiers for tasks like transaction, filing, and communication classification have beaten zero-shot prompting by wide margins, while holding steady, predictable latency regardless of task.
40+ pts
Typical accuracy gain
over zero-shot prompting.
~30ms
P50 latency, consistent
across task types.
FURTHER READING
More on training models on your own data.
GET STARTED