CTRL-0.6B

The in-browser build of the Expedient agent: a QLoRA adapter on Qwen/Qwen3-0.6B (Apache-2.0), trained on the same reviewed knowledge base as CTRL-2B.

It exists to make the demo runnable inside a browser tab: the Q4_K_M GGUF is 378 MB instead of 1.45 GB, and Qwen's architecture is supported by llama.cpp WASM builds, so CTRL-0.6B runs client-side with no server.

▶ Try it: CTRL-0.6B demo Space

This is not a replacement for CTRL-2B. Use CTRL-2B where quality matters; use CTRL-0.6B where the model has to ship to the client.

Measured results

Number-fact recall (gold numbers present in the answer) and precision (answer numbers present in gold — invented numbers are the hallucination signal). 26 held-out questions, greedy decoding.

Arm	Recall	Precision
Base Qwen3-0.6B, closed-book (26-row set)	0.692	—
CTRL-0.6B, closed-book (26-row set)	0.692	0.846
Base Qwen3-0.6B, closed-book (94-row set)	0.713	0.776
CTRL-0.6B, closed-book (94-row set)	0.713	0.776
CTRL-0.6B, open-book, retrieved excerpts (26-row set)	0.962	1.000

Retrieval hit@3: 26/26 (mean gold rank 0.0), and 94/94 on the 94-row production-shape check against the full 939-passage index.

General-capability reference (2026-11-05): 40-item probe pass rate 0.775, WikiText-2 perplexity 26.92, OOD invented-number rate 0.278.

Read that carefully. Closed-book recall is identical to the base model — at 0.6B the adapter learns voice and format but cannot store facts. All the factual reliability comes from retrieval. For comparison, CTRL-2B reaches 0.772 closed-book recall and 0.865 precision at 2.5B.

Note the encouraging part: open-book precision is 1.000 — with excerpts in front of it, CTRL-0.6B invented no numbers across the eval set.

Training

• QLoRA r16 / α32, lr 2e-4 cosine, effective batch 16, 4 epochs, 211 steps (~14 min on an RTX 4050 Laptop 6 GB), NF4 double-quant, bf16 compute
• Corpus: 939 QA rows (257 first-party reviewed + 682 promoted through automated source-page verification; see CTRL-2B for the triage method; count corrected 2026-11-05, one duplicated question row)
• Assistant-only supervision; trained format is Qwen ChatML with an empty think block: <|im_start|>assistant\n<think>\n\n</think>\n\n{answer}

Files

Path	What
ctrl-06b-v1-q4_k_m.gguf	378 MB Q4_K_M, the browser build
adapter/	LoRA weights + tokenizer
kb/	The retrieval payload the demo ships: kb_corpus.jsonl (939 passages), kb_emb.json (passage vectors), q_emb.json (query vectors for the curated questions)

Retrieval vectors are precomputed with BAAI/bge-small-en-v1.5 (CLS pooling, L2-normalised) so the browser only has to run a dot product — no embedder download and no in-browser embedding pass.

Serving

Any llama.cpp build runs the GGUF. In the browser the demo uses @wllama/wllama 2.3.1; note that wllama 2.x bundles an older llama.cpp, and Qwen-architecture GGUFs load there while some newer architectures do not.

Abstention in the demo is prompted only (an ABSTAIN instruction in the system prompt). No confidence gate is calibrated for this vehicle; the 0.74 Lens threshold documented for CTRL-2B was measured on MiniCPM5-2B and does not transfer.

Provenance

Every number on this card traces to a hash-pinned file: LINEAGE.json (training config, corpus hashes, export hashes) and MEASUREMENTS.md (every published number with its source file). The 94-row closed-book baseline and the general-capability reference numbers above were first published 2026-11-05.

Conditional-memory study (2026-11-05)

Two knowledge-injection approaches were trained, tested under the protocol above, and measured against a held-out 94-row kill bar (closed-book recall >= 0.763, i.e. +0.05 over the 0.713 baseline) plus capability gates (probe >= 95% of base, no OOD degradation):

• Engram conditional-memory sidecar (DeepSeek-style n-gram-keyed lookup table, arXiv:2601.07372), four configurations plus a first-party-only arm. Best arm gains +0.039 closed-book recall on the 26-row set and +0.012 on the 94-row set, short of the +0.05 bar, and costs 10 points of general capability; a stronger variant that memorises the corpus corrupts what the backbone already knew (precision 0.776 to 0.617, out-of-domain invented numbers 0.278 to 0.467). Full write-up, including the harness defects the study found and fixed, in ENGRAM.md. Measured; not published as a knowledge product.
• IAR retrain (CPT on the reviewed documents, QA alignment, linear recover-merge, arXiv:2608.20281): the aggressive arm reached the recall bar (+0.056) but displaced general facts (probe -10pp, WikiText-2 PPL +75%); the conservative arms held capability but gained only +0.02. No arm cleared both gates. These measurements are unaffected by the sidecar harness defects, because that study ran without the module attached.

CTRL-0.6B v1 remains the shipping model. The most promising signal for a future retrain: OOD invented-number rate improved on every IAR arm (0.278 -> 0.173-0.233), so domain-document grounding reduces hallucinated numbers even out of domain. The retrain becomes interesting when the corpus grows past this study's 73.5k tokens of reviewed text.

Governance

Base model Apache-2.0; adapter ships Apache-2.0. Corpus sources are Expedient first-party KB drafts plus vendor documentation paraphrased and source-verified against live pages. Rights in the underlying vendor content remain with the publishers; this release is intended for Expedient-internal support use.