CTRL-2B info card

01

The idea

The adapter supplies voice. Retrieval supplies truth.

CTRL-2B is a QLoRA adapter on MiniCPM5-2B, trained on 939 reviewed question and answer pairs from Expedient's cloud infrastructure knowledge base. Fine-tuning teaches tone, answer shape, and when to stop. It does not put facts in the weights.

Every supported answer runs against the top three retrieved excerpts. A lens gate scores the model's confidence in its own answer. Below 0.74, it abstains. That rule is the difference between a demo and a support agent.

01 / INPUTQuestion

02 / RETRIEVETop 3 excerpts

03 / GATELens 0.74+

AnswerAbstain

The gate is part of the runtime, not the adapter. Closed-book mode is not supported.

02

Measured

Number-fact recall counts gold numbers present in the answer. Precision counts answer numbers present in the gold answer, so invented specifics lower the score.

Arm	Recall	Precision
Base MiniCPM5-2B, closed-book	0.692	—
CTRL-2B, closed-book	0.772	0.865
CTRL-2B, open-book, production shape	0.943	0.945

Open-book production shape: 94 held-out rows, 939-row index, retrieval hit@3 94/94 at rank 0.0. The 26-row probe scored 0.962 / 0.962. GGUF Q8_0 engine eval: 1.000 / 0.981 open-book, overall 0.889.

OOD max0.723

Gate0.740

Paraphrase p250.748

v2 robustness probe: 26 paraphrases plus 15 out-of-domain questions. The gate sits between the highest OOD confidence and the lower quartile of in-domain paraphrases. The 0.74 threshold was calibrated on the v1 41-row probe.

03

Corpus

Five sources feed the v2 corpus. The belt shows rows collected before review; the gate decides what ships. Select a plate to see the source.

BELT LIVE 5 SOURCES / 1,377 ROWS

01 / 05VENDOR DOCS

Nutanix

522 source rows · nutanix-knowledge

Reviewed rows enter the v2 corpus; the remainder waits for review.

Five source sets total 1,377 rows. The v2 corpus holds 939 after review and dedupe. The 1,005 unreviewed vendor rows triaged to 682 accepted, 207 rejected, and 116 held for human review.

Source	Type	Rows
Nutanix	Vendor docs	522
VMware VCF9	Vendor docs	504
Expedient KB	First-party	218
Zerto	Vendor docs	71
Cohesity	Vendor docs	62
Five source sets		1,377

Nutanix and VMware marks via Simple Icons. Zerto, Cohesity, and Expedient plates are typographic placeholders. Vendor names and marks remain trademarks of their owners.

04

Quick start

The adapter ships with the GGUF Q8_0 build. The runtime that adds retrieval and the gate lives in the training repo. This card covers the model artifact.

bash / llama.cpp

# GGUF Q8_0, 2.5 GB
llama-server -m ctrl-2b-q8_0.gguf -c 4096

python / peft

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("openbmb/MiniCPM5-2B")
model = PeftModel.from_pretrained(base, "chadstthomas/ctrl-2b")
tok = AutoTokenizer.from_pretrained("openbmb/MiniCPM5-2B")

Closed-book generation is not a supported posture. Pair the adapter with retrieval and the lens gate.

05

Limits

01

Facts live in retrieval.

Closed-book precision is 0.865, not 1.0. Specific numbers, versions, and dates need an excerpt behind them.

02

The gate is mandatory.

Prompted refusal caught 20% of out-of-domain questions. The lens gate caught 15/15 at 0.74. Do not ship the adapter without it.

03

Scope is narrow.

Expedient cloud infrastructure support. It is not a general assistant, a coding model, or a substitute for the KB owner's review.

The 0.74 threshold is calibrated for this vehicle. Recalibrate after a base-model, corpus, or adapter change.

06

Training

Base model: MiniCPM5-2B (2.52B)
Method: QLoRA, r16 / a32
Corpus: 939 reviewed QA rows
Schedule: 4 epochs, 211 steps
Hardware: RTX 4050 Laptop, 6 GB
Wall time: 54 minutes
Compute: bf16, NF4 double quant
License: Apache-2.0

A 4B control run bought no clear quality gain for twice the size, at 19.9 seconds per step versus 15.4. 2B ships.

07

Governance

The base model and this adapter ship under Apache-2.0. The corpus combines Expedient first-party KB drafts with vendor documentation, paraphrased and verified against live pages. Rights for the underlying vendor content remain with the publishers. This release is intended for Expedient-internal support use pending the corpus owner's final call.

Known limits: facts live in retrieval, not in the weights. The abstention gate is mandatory in production.

citation / bibtex

@misc{ctrl2b2026,
  title  = {CTRL-2B: A retrieval-grounded support agent},
  author = {St. Thomas, Chad},
  year   = {2026},
  url    = {https://huggingface.co/chadstthomas/ctrl-2b}
}