A 32B model that knows when not to refund
GRPO with a refund-policy reward on 180k resolved tickets. Replaced a prompt-chained frontier model.
+19 pts resolution · −81% cost
New GRPO with custom Python verifiers →
Post-training, for everyone.
POFE is the workspace where a team takes an open model and makes it great at one job. Curate the data, run SFT, DPO or GRPO, check every result against evals you trust, and ship the endpoint. No platform team required.
No signup for the demo·Private workspace on request·AWS, GCP, Azure, Databricks, on-prem
acme-support
Datasets 6
Reward functions 3
Runs 27
Eval suites 4
Endpoints 2
Recent runs
grpo-07 ●
dpo-04 0.861
sft-12 0.813
sft-11 0.790
Qwen/Qwen3-32B · GRPO · LoRA r=64
policy lossmean rewardKL 0.042ETA 1h 12m
Eval vs. base
resolution0.710.88
policy0.820.97
tone (judge)3.94.6
refusal-safety0.990.99
latency p50410ms430ms
GPU utilisation
gpu0–3 · 94%
gpu4–7 · 91%
rollout workers · 78%
Runs where your data already lives
AWSGoogle CloudAzureDatabricksSnowflakeKubernetesSlurmAir-gapped
Five tools you were going to glue together. One workspace.
Every dataset, run, eval and endpoint is linked in one lineage graph, so any score on a dashboard traces back to the data and code that produced it.
data/tickets-v4 · pipeline184,212 rows
01 sources3://acme-logs/support/2025-*1.24M rows
02 filterresolved = true AND csat ≥ 4412k
03 redactPII: names, emails, card numbers, IBAN37,804 spans
04 dedupeMinHash, Jaccard ≥ 0.85−211k
05 synthesiseteacher → 2 candidates, judge picks preferred+48k pairs
06 splittrain 95% / eval 5%, stratified by intent184k
recipes · choose a method
| 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× |
evals/compare · release-v2n = 2,400
| 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 |
gate: must beat prod on every slice ±CI → grpo-07 passes
broker · 8× H100 · spot preferredcapacity, live
neocloud-a · us-east64 freeselected
your AWS reservation · us-west-232 freefallback
hyperscaler-b · eu-west128 freestandby
neocloud-c · eu-north16 freestandby
hyperscaler-a · eu-central0 freeno capacity
checkpoint every 15 min · auto-resume on preemption · est. 165 GPU-hours
# 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)
Every run, side by side.
Runs, not slides. Method, data version, eval delta and cost on one row, for every experiment your team has ever launched.
27 runs · acme workspace
| Run | Base model | Method | Dataset | Eval Δ vs base | Curve | Compute | Status |
|---|---|---|---|---|---|---|---|
| support-agent-grpo-07 | Qwen/Qwen3-32B | GRPO | tickets-v4 | +17.4 pts | 165 GPU-h | training | |
| claims-extract-dpo-04 | meta-llama/Llama-3.3-70B-Instruct | DPO | claims-v7 | +9.1 pts | 118 GPU-h | done | |
| sql-copilot-grpo-12 | Qwen/Qwen3-8B | GRPO | sql-verifier-v2 | +21.8 pts | 39 GPU-h | done | |
| icd10-coder-sft-03 | google/gemma-4-31B-it | SFT | notes-deid-v5 | +14.2 pts | 23 GPU-h | done | |
| brand-voice-kto-02 | zai-org/GLM-5.3-Flash | KTO | feedback-q3 | +6.3 pts | 14 GPU-h | done | |
| kb-embed-ft-02 | nvidia/Nemotron-3.5-Embed-1B-BF16 | SFT | kb-pairs-v3 | +6.8 pts | 6 GPU-h | done | |
| router-distill-01 | Qwen/Qwen3-8B ← deepseek-ai/DeepSeek-V4-Flash | SFT | router-traces | n/a | queued | est. 56 GPU-h | queued |
14k+post-training runs launched on POFE last quarter
11×median inference cost cut after distilling to a tuned 8B
41 minmedian dataset-to-first-eval on managed GPUs
6 daysmedian first call to model in production, managed engagements
Training lives in the hours inference leaves idle.
On your own GPUs or reserved capacity, inference load follows your users' day. Post-training is batch work: preemptible, checkpointed every 15 minutes, happy to run at 03:00. Scheduled into the quiet hours, idle GPU-hours become models.
Yields to servingJobs give the GPU back within one checkpoint interval when traffic rises.
Per-pool policyPriority, region and GPU class set per pool and per project.
MeteredGPU-hours reported per project and per run, next to tokens served.
| Hour | Average inference load | Nodes training |
|---|---|---|
| 00:00 | 13% | 10 of 10 |
| 01:00 | 10% | 10 of 10 |
| 02:00 | 9% | 10 of 10 |
| 03:00 | 8% | 10 of 10 |
| 04:00 | 15% | 10 of 10 |
| 05:00 | 21% | 10 of 10 |
| 06:00 | 27% | 10 of 10 |
| 07:00 | 38% | 7 of 10 |
| 08:00 | 50% | 0 of 10 |
| 09:00 | 58% | 0 of 10 |
| 10:00 | 67% | 0 of 10 |
| 11:00 | 77% | 0 of 10 |
| 12:00 | 85% | 0 of 10 |
| 13:00 | 86% | 0 of 10 |
| 14:00 | 90% | 0 of 10 |
| 15:00 | 88% | 0 of 10 |
| 16:00 | 84% | 0 of 10 |
| 17:00 | 77% | 0 of 10 |
| 18:00 | 70% | 0 of 10 |
| 19:00 | 60% | 0 of 10 |
| 20:00 | 49% | 1 of 10 |
| 21:00 | 40% | 5 of 10 |
| 22:00 | 29% | 10 of 10 |
| 23:00 | 20% | 10 of 10 |
Click it, or script it. Then ship it.
The workspace and the SDK are the same system: researchers live in Python, product teams in the UI, and both see the same runs, evals and lineage. A gated run deploys to wherever you serve, and production feeds the next run.
- Your datalogs, tickets, docs
- Post-trainingSFT · DPO · GRPO
- Eval gatespaired CIs, calibrated judges
- Your endpointany OpenAI-compatible server, or your inference provider
- Productiontokens, latency, feedback
- Back to trainingopt-in feedback from your own request logs
- Recipes as code, checked into your repo, reproducible to the seed
- Reward functions in plain Python, sandboxed and unit-tested
- Attach a notebook to a live run and read rollouts mid-training
- Diff any two runs: data, hyperparameters, scores, sample outputs
- Deploy to a POFE endpoint, your own vLLM or SGLang servers, or your inference provider
import pofe ws = pofe.workspace("acme-support") ds = ws.datasets.get("tickets-v4").split(eval=0.05) # reward: resolved, on-policy, no invented refunds @pofe.reward def resolution(sample, completion): r = ws.judges["resolution@v3"].score(sample, completion) if pofe.checks.mentions_refund(completion) and not sample.refund_ok: return -1.0 return r run = ws.train( base="Qwen/Qwen3-32B", method="grpo", adapter="lora:64", data=ds.train, rewards=[resolution], kl=0.05, compute=pofe.Compute(gpu="h100", count=8, max_price=2.60), ) run.gate(ws.evals["release-v2"], must_beat="prod")
# a deploy target is any OpenAI-compatible server that can load an adapter target = pofe.targets.OpenAICompatible( endpoint="https://<your-inference-endpoint>/v1", runtime="vllm", # or "sglang", a POFE endpoint, or your provider ) dep = run.deploy(target=target, model_id="acme/support-agent", canary=0.10)
from openai import OpenAI client = OpenAI(base_url="https://<your-inference-endpoint>/v1", api_key=os.environ["INFERENCE_API_KEY"]) r = client.chat.completions.create( model="acme/support-agent", # the tuned adapter, called like any model messages=[{"role": "user", "content": "I was charged twice"}], )
# your logs, your bucket: nothing leaves your account logs = ws.connectors.request_logs("s3://acme-logs/inference/") fresh = logs.since("2026-10-01").where(outcome="resolved").redact_pii() ws.datasets["tickets-v5"] = ws.datasets["tickets-v4"].extend(fresh).dedupe() ws.schedule(run.recipe, data="tickets-v5", every="30d")
384 custom models. 1.2T tokens a month.
Month 24 for a team shipping custom models on open base models, at assumed market rates. The sliders are yours.
- Custom models live
- Tokens served per month
- GPU-hours per month
- Serving cost per month
Assumed market rates: serverless tokens by model class, dedicated at $3.00/H100-hour, training at $5.00/B200-hour. Not POFE prices. Method · Economics of custom models.
Scores you can defend in a review.
Fine-tuned models routinely look better than they are because the eval leaked or a judge agreed with itself. POFE tests the evals too.
- Judge agreement with your human labels, per rubric, as Cohen's κ
- Contamination scans between every training set and every eval
- Paired confidence intervals, so a +0.4 point "win" is called noise
- Release gates that block a deploy when any slice regresses
eval-health · release-v2412 human labels
| rubric | judge κ | leak | 95% CI | status |
|---|---|---|---|---|
| resolution | 0.81 | 0.0% | ±1.2 | trusted |
| policy-adherence | 0.86 | 0.0% | ±0.9 | trusted |
| tone | 0.47 | 0.0% | ±3.8 | recalibrate |
| kb-grounding | 0.79 | 6.4% | ±1.6 | contaminated |
| refusal-safety | 0.93 | 0.0% | ±0.3 | trusted |
→ 312 kb-grounding items overlap the training set (13-gram). Excluded from gate.
→ tone judge disagrees with reviewers on sarcasm. Suggest 40 more labels.
→ tone judge disagrees with reviewers on sarcasm. Suggest 40 more labels.
“We had three teams fine-tuning three different ways and no way to compare them. Six weeks after moving to POFE we retired our largest API contract, and the tuned model beats it on every slice we care about.”
VP of Machine Learning · Fortune 100 insurer
−72%inference spend on claims triage
0.94F1 on policy-clause extraction, up from 0.78
VPCruns inside their own account; no data egress
ICD-10 coding that clinicians sign off
SFT on de-identified notes plus DPO from coder corrections, on Databricks next to the source data.
0.91 exact matchClause extraction at a global law firm
Managed engagement: eval design, synthetic hard cases, and an 8B distilled model deployed on-prem.
4.1× faster reviewDo it yourself, or have it done.
Same platform underneath. Start self-serve and call us for the hard part, or start managed and take the keys when your team is ready.
Self-serve
The platform
For engineers who know what they want to train. Bring a model and a JSONL, and press run.
- StartPick a base model, upload a JSONL, press run
- RunOn our GPUs, your cloud account or Databricks
- OwnDownload the weights whenever you like
Managed
The team
For organisations that need the outcome, not the tooling. Our research engineers own it end to end.
- ScopeTarget metric agreed in week one
- BuildEvals, data, training, distillation
- Hand overModel, pipeline and runbook in your account
From the research notebook.
Reproducible experiments, plus Post-training algorithms 101: one page per algorithm, from SFT to GRPO.
Simulation
GRPO group size under a fixed rollout budget
Why most groups carry no gradient, and when smaller groups win.
Read → StatisticsHow many eval examples do you need?
Power curves for 1–8 point gains, paired vs unpaired.
Read → Cost modelWhen a tuned model beats the API on cost
A break-even model and which assumptions actually move it.
Read →Questions people ask first.
Something else? Ask an engineer and you'll get a technical answer, not a demo request.
What is post-training?
Post-training is everything that happens to a model after pre-training: supervised fine-tuning (SFT) on examples of the behaviour you want, preference optimization such as DPO, and reinforcement learning such as GRPO or PPO against reward models or verifiable checks. It is how a general open model becomes a specialist that follows your formats, policies and tone. Our complete guide covers it end to end.
Which models can I post-train on POFE?
More than 200 open-weight base and instruct models, from 8B-class models such as Qwen/Qwen3-8B and meta-llama/Llama-3.1-8B-Instruct to large mixture-of-experts models such as deepseek-ai/DeepSeek-V4-Pro, moonshotai/Kimi-K3 and zai-org/GLM-5.3, plus embedding, reranking and transcription models. Large MoE models are trained as LoRA adapters on multi-GPU nodes. You can also bring your own checkpoint from Hugging Face, S3, GCS, Azure Blob or a Unity Catalog volume.
Do I need my own GPUs?
No. POFE sources H100, H200 and B200 capacity across a network of clouds and neoclouds and schedules your run where it is cheapest and available. If you already have reserved capacity, a Kubernetes or Slurm cluster, or Databricks GPU compute, POFE runs there instead.
Does my data leave my environment?
Not if you don't want it to. In VPC and on-prem deployments the control plane only sees run metadata; datasets, checkpoints and logs stay in your buckets and your accounts.
More questions (3)
How is POFE different from a fine-tuning API?
A fine-tuning API gives you a job endpoint. POFE gives you the whole loop: data curation and synthesis, every major training method, eval suites with judge calibration, GPU sourcing, and inference deployment, all versioned together so you can see exactly which data and which run produced which score. And you can take the weights with you.
Where do trained models run?
Wherever you serve today: a POFE endpoint, your own vLLM or SGLang servers, or your inference provider through its OpenAI-compatible API and adapter upload. The weights are yours either way.
Can POFE do the post-training for us?
Yes. Managed post-training is a white-glove engagement where our research engineers own the outcome: they design the evals, build the data, run the experiments and hand you a model that clears agreed metrics, plus the full pipeline so your team can keep iterating.
Open the workspace. Train something.
The demo runs in your browser with sample data and takes about three minutes. When you're ready for your own data, an engineer will set up a private workspace with a baseline eval.