Dependabot for fine-tuned models
Your model, requalified on every new base.
A new base ships. Roadbed retrains your model on your GPUs, scores it on evals from your own reviewers, and hands your risk team the evidence.
Qwen3.8-27B shipped. Candidate ready for claims-triage.
v12 on Qwen3.6-27B → v13 on Qwen3.8-27B · 38,410 verdicts · 1,284 frozen holdout cases · material change
- Your evals
- +2.1 94.0 → 96.1 · 95% CI +0.9 to +3.3
- Cost
- −31% per 1k calls, your GPUs
- Below margin
- 0 of 14 slices, margin −2.0 · 2 flagged low-n
- License
- Allowed Apache-2.0 · policy allow-list · BOM attached
How it works
Capture. Rebase. Promote.
The system of record for what your models have learned, and the control plane for changing them.
No platform we know of requalifies a fine-tuned model per slice when a new base ships, or writes the change record a risk team reviews. That loop is Roadbed.
- Open-weight bases ship every few weeks: Qwen3.8, DeepSeek V4.1, Mistral Large 4.
- Closed fine-tuning APIs retire models on notice.
- Regulators ask for a change record on every update.
Capture anywhere, even on a closed API. Rebase onto open weights when you are ready.
-
01
Capture verdicts
Every accept, edit and rejection, logged with who and why.
-
02
Derive evals
Golden sets, correction tests and slices. A sealed holdout.
-
03 The gap
Rebase on every new base
Retrained and scored on your evals, on your hardware.
-
04
Promote behind a gate
Shadow, canary, rollback. An evidence pack for your validation file.
Insurance. A claims-triage model on a closed API gets a retirement date.
-
01verdictClaim 48213 · fast-track → complexreason: injury noted in FNOL · adjuster, 09:14
Capture adjuster verdicts
Accepted, re-routed or escalated, with the reason.
-
02evalsSlice: auto · line: bodily injurygolden 412 · correction tests 96 · coverage 93%
Derive evals by line
Golden claims by line and state. A sealed holdout.
-
03 The gaprebaseQwen3.8-27B → candidate v13your evals +2.1 (CI +0.9 to +3.3) · cost −31%
Rebase the day the notice lands
A scored candidate in about a day. You decide when it ships.
-
04gateShadow 100% · canary 10%0 slices below margin · bias tests pass · awaiting MRM sign-off
Promote with the evidence attached
Shadow first, canary on one line. Evidence pack attached.
Banking. A KYC model ships on overall accuracy. Portuguese ID cards drop 9 points, and nobody notices for a quarter.
-
01verdictField: date of birth · correcteddoc: PT identity card · analyst, 11:02
Capture analyst verdicts
Every field accepted or corrected on an ID.
-
02evalsSlice: PT identity cardsgolden 188 · coverage 61% → 97%
Derive evals per slice
Document type, issuing country, language. Each slice gets a number.
-
03 The gaprebaseCandidate v13 · scored per slicePT identity cards −9.0 · overall +1.8
Rebase and score per slice
A 9-point drop on one slice shows before anything ships.
-
04gatePromotion blockedreason: slice below margin · finding logged
Promote only past the gate
Blocked. The finding is logged, with the evidence your validators ask for.
Healthcare. A clinical coding model cannot leave the network, so it stays on Llama 3.1 while newer bases pass it.
-
01verdictCode I21.4 → I21.9 · changeddoc: discharge summary · coder, ward 7
Capture coder verdicts
Every code confirmed or changed, inside your network.
-
02evalsSlice: cardiology · summariesgolden 240 · holdout sealed · on your disks
Derive evals on your disks
By specialty and document type, where the records live.
-
03 The gaprebaseLlama 3.1 → Gemma 4 12Badapter fine-tune in your cluster · nothing left the network
Rebase on your GPUs
A Helm chart in your cluster. Nothing leaves.
-
04gateApprovals 2 of 2 · BOM attachedrollback: one click
Promote under your governance
Approvals, lineage, bill of materials. One-click rollback.
Agents. A coding agent tuned on the monorepo last spring. Qwen3-Coder-Next ships; the agent stays put.
-
01verdictPR #2231 · mergedtests 212 passed · review: approved
Capture engineering verdicts
Merged changes, passed tests, review comments.
-
02evalsEvals from your repo · 1,840 casesheld-out PRs · replayed reverts
Derive evals from your repo
Held-out pull requests and replayed reverts.
-
03 The gaprebaseQwen3-Coder-Next · scored on your repoheld-out PRs: tests pass 71% → 75% · 0 reverts regress
Rebase on every coder base
Retrained on your history, scored on held-out PRs.
-
04gateCanary: platform teamrollback if merge rate drops
Promote to a canary team
One team first. Roll back if merge rate drops.
Composite scenarios from real demand. Numbers illustrative.
What you own
Plain files in your bucket.
The asset is a folder, not a platform. Open format and CLI on GitHub when the first design partner is live.
model: claims-triage
base: Qwen/Qwen3.8-27B@7f3a…
data:
verdicts: s3://acme/claims/verdicts@2026-10-28
holdout: sealed@2026-01-14
evals:
golden: 412
correction_tests: 96
slices: [line, state, channel]
margin: -2.0
recipe: lora-r16-sft@v4
gate: [shadow, canary, bias-tests, sign-off]
lineage: v12 ← Qwen3.6-27B · verdicts@2026-07-02
$ roadbed watch --base Qwen/Qwen3.*
new base: Qwen/Qwen3.8-27B
$ roadbed rebase claims-triage --on Qwen/Qwen3.8-27B
trained on 38,410 verdicts
scored on 1,284 holdout cases
+2.1 (CI +0.9 to +3.3)
0 of 14 slices below margin
opened rebase #13 → awaiting MRM sign-off
The rebase moment
A pull request against your model.
Retrained on your data, scored on your evals, costed on your hardware. When a slice fails, the gate says no.
- New base detected: Qwen/Qwen3.8-27B
- Candidate v13 trained on 61,200 verdicts
- Scored on 1,420 sealed holdout cases. Overall +1.8
- PT identity cards −9.0, below the −2.0 margin
- Promotion blocked. Finding logged against the model
- v12 stays live. Evidence pack attached for review
Where a wrong answer costs money.
Insurance
Claims triage that survives the next base.
A retirement notice becomes a scored candidate in a day.
See the loopBanking
KYC documents, checked slice by slice.
A 9-point drop on one issuing country is caught before it ships.
See the loopHealthcare
Clinical coding that never leaves the network.
A Helm chart in your cluster, training on your GPUs.
See the loopCoding agents
An agent that keeps up with coder bases.
Every merged change is a verdict. Qwen3-Coder-Next becomes a candidate scored on your repo.
See the loopSpeech
Whisper or Nemotron Transcribe, rebased on every release.
Verdict: word error rate on your frozen holdout, plus every human fix.
Same loop, different verdictRobotics
VLA policies like π0 that outlive their backbone.
Verdict: success rate in simulation and on the cell. Trigger: a new VLM backbone.
Same loop, different verdictEvery model runs on your roadbed.
Read the manifestoFine-tuned a model?
Three design-partner slots for teams whose fine-tuned model sits on an API that is being retired, or on a base that has fallen behind. Or leave an email for one update.
Design partners and investors: founders@roadbed.ai