Hugging Face's logo
Activity Feed

AI & ML interests

None defined yet.

Recent Activity
chadstthomas  updated a dataset 1 day ago
expedient-ai/vmware-vcf9-knowledge
chadstthomas  updated a dataset 2 days ago
expedient-ai/nutanix-knowledge
chadstthomas  updated a dataset 2 days ago
expedient-ai/cohesity-knowledge
View all activity

Organization Card
Community About org cards

Cloud + Data + AI

Expedient AI CTRL Models

The adapter supplies the voice. Retrieval supplies the truth.

94/94
Retrieval hit@3
0.962
Open-book recall
1.000
Number precision
378 MB
Browser build
Try the demo How CTRL-2B works

Expedient is a Cloud Service Provider delivering Intelligent Infrastructure for the AI era. Twenty-five years of service. 99% client retention. Every solution backed by a 100% SLA. These models answer support questions from Expedient's own knowledge base, and every card states what was measured and where the model stops being reliable.

How an answer is produced

1
Question
A support question about Expedient or the platforms it operates.
2
Retrieve
bge-small retrieval over 939 reviewed KB pairs returns the top 3 excerpts.
3
Gate
Lens confidence at 0.74 or above: answer. Below: abstain and point to the KB.

Retrieval supplies the facts. The adapter sets the format. Closed-book generation is not a supported posture, and the gate is part of the runtime, not the adapter.

Measured

ModelBaseDownloadClosed-bookOpen-bookPrecision
CTRL-2B MiniCPM5-2B2.68 GB 0.7720.9620.945
CTRL-0.6B Qwen3-0.6B378 MB 0.6920.9621.000
CTRL-2B-ONNX CTRL-2B merged1.62 GB n/en/en/e

Recall counts the numbers in a gold answer that also appear in the model's answer, on 26 held-out questions. Precision counts answer numbers that exist in the gold answer, so invented specifics lower it.

CTRL-0.6B runs the demo inside a browser tab: 378 MB, cached after the first visit, with number precision of 1.000 open-book on the eval set.

What it does well, and what it does not

  • KB-grounded answers in the Expedient voice
  • Abstains on out-of-domain questions at the measured gate
  • Every number traces to a run artifact
  • Closed-book mode is not supported, at either size
  • Scope is Expedient cloud infrastructure support
  • The 0.74 gate is calibrated for this vehicle only

Corpus

Every row enters training only after review. Of 1,005 unreviewed rows, 682 were promoted after checking each answer against its live source page, 207 were rejected for invented figures or missing entities, and 116 were held for a person to read. Unreviewed imports stay out.

DatasetRowsContents
expedient-engram 1,417Master collection: first-party KB, company facts, source-verified vendor documentation
nutanix-knowledge 522Nutanix
vmware-vcf9-knowledge 504VMware Cloud Foundation 9
expedient-kb-knowledge 218Expedient KB articles
zerto-knowledge 71Zerto
cohesity-knowledge 62Cohesity

Go deeper

The CTRL-2B walk-through, with the gate calibration and the training record, lives at expedient-ai/ctrl-2b-info. The demo runs at expedient-ai/ctrl-0.6b-demo.

Adapters ship under Apache-2.0, matching their Apache-2.0 bases: MiniCPM5-2B from OpenBMB and Qwen3-0.6B from Alibaba. Rights in vendor documentation remain with the publishers. These releases support Expedient's own support operations.
measurements: runs/qwen3-06b-qlora-v1 · runs/minicpm5-2b-qlora-v2

Collections 2