Track B · XRFM CoderSec

XRFM CoderSec

A defensive-coding research line based on an explicitly attributed pretrained model. Adapter and merged weights are not publicly available.

Model lineage

CoderSec is not an XRFM Native architecture and was not trained from random initialization. It uses Qwen/Qwen3-1.7B, revision 70d244cc86ccca08cf5af4e1e306ecf908b1ad5e, under Apache-2.0, followed by NF4 + FP16 QLoRA fine-tuning in the project's training scripts.

Dataset

Training used xrfm-codersec-sft-40k revision 808d7505. The published dataset has 24,100 rows despite “40k” in its name. The audit found malformed ChatML in every row and placeholder code in all 2,100 code examples; see the datasets page.

Training record

Repository experiment records describe base-model bake-off, multiple ablations and a flagship QLoRA run on Tesla T4 hardware. The base bake-off preferred Qwen3-1.7B over Qwen2.5-Coder-1.5B-Instruct on a small internal benchmark. These are experiment records, not independent benchmarks.

Evaluation

Public documentation describes an 8-prompt weaponization screen, 7 false-refusal prompts and 4 Roman-Urdu perturbation pairs. These are preliminary, limited-sample internal checks. A zero observed failure rate on a small sample does not establish zero real-world risk.