The whole post-training loop, in one place.
Data, training, evals, compute and inference used to be five tools and three teams. In POFE they are five tabs of the same workspace, sharing one lineage graph — so every score traces back to the data and run that earned it.
Turn production exhaust into training data.
Point POFE at logs, tickets, transcripts or documents in your bucket, warehouse or Unity Catalog. Build a pipeline once; re-run it every week as new data lands.
- Connectors for S3, GCS, Azure Blob, Snowflake, BigQuery, Delta and Unity Catalog
- PII detection and redaction with reviewable spans
- MinHash de-duplication and quality filters
- Synthetic SFT examples and preference pairs from any teacher model
- Human labelling queues for the 2% of data that actually needs a person
Every method that matters, tuned per model family.
Pick a base, a method and an adapter. POFE ships recipes that are already tuned for each family's chat template, tokenizer quirks and stable learning rates, so your first run is a real baseline, not a debugging session.
- SFT, DPO, ORPO, KTO, SimPO, PPO, GRPO and on-policy distillation
- LoRA, QLoRA, DoRA or full-parameter; FSDP and DeepSpeed ZeRO-3 multi-node
- Reward functions in Python, sandboxed, with unit tests
- Live rollouts: read what the policy is generating while it trains
- Bring your own container — TRL, Axolotl, torchtune, verl, NeMo
| method | needs | best for |
|---|---|---|
| SFT | demonstrations | format, style, domain |
| DPO / ORPO | chosen vs rejected | tone, preferences |
| KTO | thumbs up / down | product feedback logs |
| GRPO | reward fn or verifier | reasoning, tool use, agents |
| PPO | reward model | open-ended RLHF |
| Distill | teacher model | shrinking cost 5–20× |
Ship on evidence, not on vibes.
Evaluation is where most post-training projects quietly fail. POFE makes evals reproducible, calibrated and slice-aware, and wires them into release gates.
- Golden sets with versioning and owner sign-off
- LLM-as-judge with agreement scores against your human labels
- Contamination scans between every training set and every eval
- Per-slice comparisons with bootstrap confidence intervals
- Gates that block a deploy when any slice regresses
| slice | prod | sft-12 | grpo-07 |
|---|---|---|---|
| billing | .74 | .81 | .90 |
| refunds | .69 | .77 | .92 |
| account access | .83 | .85 | .87 |
| escalations | .61 | .66 | .79 |
| non-English | .72 | .70 | .74 |
We find the GPUs. You set the ceiling.
The cheapest H100 hour is rarely in the account you already have. POFE's broker quotes capacity across clouds, neoclouds and your own reservations, then places, checkpoints and resumes the job for you.
- H100, H200, B200 and MI300X capacity across a multi-provider network
- Price caps, region pinning and data-residency constraints
- Spot with automatic checkpoint-resume — preemptions cost minutes, not runs
- Use your own Kubernetes, Slurm or Databricks GPU clusters at no markup
- One invoice, or charge back to your cloud commitments
From checkpoint to endpoint in one click.
Deploy a run to an autoscaling, OpenAI-compatible endpoint, or export it anywhere. Production traffic flows back into the data studio, closing the loop.
- vLLM and SGLang runtimes with speculative decoding and FP8
- Multi-LoRA: hundreds of adapters share one base model's GPUs
- Canary and shadow traffic with live eval scoring
- Export to Databricks Model Serving, SageMaker, Vertex AI, Azure ML or Kubernetes
- Weights are yours: download safetensors at any time
# OpenAI-compatible. Swap the base URL, keep your code. curl https://api.pofe.ai/v1/chat/completions \ -H "Authorization: Bearer $POFE_KEY" \ -d '{ "model": "acme/support-agent@grpo-07", "messages": [{"role":"user","content":"Card charged twice"}] }' # endpoint: support-agent # replicas 2→9 (autoscale) · p50 212ms · p99 640ms # traffic grpo-07 10% · sft-12 90% (canary) # adapters 41 on one base (multi-LoRA)
Your data stays where you put it.
Three deployment models, one product. The control plane never needs your data; the data plane runs where your security team says it runs.
| POFE Cloud | Your VPC | On-prem / air-gapped | |
|---|---|---|---|
| Where training runs | POFE-sourced GPUs | Your AWS, GCP, Azure or Databricks account | Your data centre |
| Where data lives | Isolated, encrypted per-tenant storage | Your buckets and catalogs | Your storage |
| Identity | SSO, SCIM | SSO, SCIM, your IAM roles | SSO via your IdP, offline licence |
| Audit | Full audit log export | Logs to your SIEM | Logs to your SIEM |
| Best for | Speed, startups, pilots | Most enterprises | Regulated and sovereign workloads |
The details.
Deeper dives live in the guides, starting with choosing between SFT, DPO and GRPO.
Which training methods does POFE support?
Supervised fine-tuning (SFT), continued pre-training, DPO, IPO, ORPO, KTO, SimPO, reward model training, PPO, GRPO and DAPO-style RL with verifiable rewards, and on-policy distillation from a teacher model. Each method works with LoRA, QLoRA, DoRA or full-parameter training, single-GPU to multi-node.
Can I bring my own training code?
Yes. Recipes are Python. You can use ours, modify them, or submit your own container that uses TRL, Axolotl, torchtune, NeMo, verl or plain PyTorch. POFE still handles data mounting, GPUs, checkpoints, metrics, evals and lineage.
How does GPU sourcing work?
You set a GPU type, count, region constraints and a maximum hourly price. POFE's broker places the job on the cheapest available capacity across its supply network or your own accounts, checkpoints frequently, and resumes automatically if a spot node is reclaimed.
What does inference deployment include?
Autoscaling endpoints on vLLM or SGLang with an OpenAI-compatible API, multi-LoRA serving so hundreds of adapters share one base model, scale-to-zero, shadow and canary traffic, and request logging that feeds straight back into your datasets. You can also export to Databricks Model Serving, SageMaker, Vertex AI or your own Kubernetes.
See it on your data.
A 45-minute technical walkthrough with a post-training engineer, using a sample of your own data if you bring one.