Test the support policy
Measure exact output format and routing decisions. Make a failed acceptance check visible instead of relying on a lower training loss.
Forgedata’s initial product focus is the engineering team adapting an open model to its SaaS support workflow. The first local reference covers structured ticket triage. Customer demand and production value remain hypotheses to validate.
The team has approved, resolved tickets and needs consistent routing, escalation, or structured outputs inside its existing support application.
This is an initial customer profile, not a claim of current customers, interviews, or pilot commitments.
Discuss your support workflow“Someone we do not recognise downloaded our customer information.”
A valid JSON response can still make the wrong decision. In the local reference run, the adapter classified this synthetic security ticket as an access problem. Evaluation exposed that failure and blocked promotion.
Inspect the actual failed exampleThe reference uses Transformers and PEFT. The product direction is the support-specific preparation, policy tests, and traceability around those tools.
Measure exact output format and routing decisions. Make a failed acceptance check visible instead of relying on a lower training loss.
Keep dataset fingerprints, a pinned base model, training settings, evaluation outputs, and saved-adapter checks together.
Use standard PEFT adapters and reproducible instructions. Target deployment compatibility must still be tested for each model and runtime.
These distinctions describe the current work, so a software roadmap is not mistaken for an operational SaaS backend.
| Capability | Current evidence | Remaining product work |
|---|---|---|
| Dataset preparation | Synthetic JSONL, schema checks, exact deduplication, fixed splits | Customer data permissions, import UX, version management |
| LoRA adaptation | Completed local run, saved adapter, reload verification | Hosted job orchestration, access control, NVIDIA execution |
| Policy evaluation | All baseline/candidate predictions and a failed promotion gate | Customer-defined acceptance criteria and broader evaluation |
| Deployment | Standard local adapter artifact | Hosted inference endpoint, monitoring, licence and compatibility review |
If a prompt already meets quality, cost, and maintenance requirements, a training run may add unnecessary complexity.
Current product documentation and account facts belong in an authorised retrieval or application layer. Fine-tuning does not replace current knowledge or access control.
The intended Forgedata contribution is coordination and repeatable policy evaluation. We have not demonstrated superiority over competing training platforms.
We plan to test this direction with SaaS teams before claiming operational impact or paid demand.
Bring a use case. Explore a model workflow built for your team.
Request access