ab9b-v2

Defensive-cyber model on Qwen3.5-9B. Merged and quantised to Q4_K_M. Run the GGUF. The LoRA is here if you want the bf16 base plus the adapter instead.

Trained for CVE→CWE, CVSS vectors, threat-intel questions, vulnerability review, and blue-team / DFIR / audit loops that call tools. Answers are short and lead with the conclusion. Thinking is off.

Run

huggingface-cli download dorb/ab9b-v2 ab9b-v2-Q4_K_M.gguf --local-dir .

llama-server -m ab9b-v2-Q4_K_M.gguf -ngl 999 -fa on --jinja \
  -c 32768 -np 1 --chat-template-kwargs '{"enable_thinking":false}'

OpenAI tools work with --jinja. The text protocol it was trained on is CALL: tool_name(arg="…") and then FINAL: ….

The LoRA that was merged into that GGUF:

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "Qwen/Qwen3.5-9B"
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(model, "dorb/ab9b-v2", subfolder="adapters/v2a-ck900")
tokenizer = AutoTokenizer.from_pretrained(base)

adapters/v2b-final and adapters/v2c-final are the later rounds. They were scored on the dev tier and not selected. Details are in evidence/REPORT.md.

Test tier

One scoring pass. Reference is ab9b-r1 on the same base. Paired bootstrap, 95% CI.

	ab9b-v2	ab9b-r1	base
cyber composite, 7 families	0.739	0.725	0.719
CVE → CWE (n = 329)	0.763	0.608
CVSS vector (n = 334)	0.812	0.715
native tool-call loop, solved (n = 87)	0.483	0.172	0.157
plain CALL: loop, solved (n = 89)	0.528	0.483	0.236
single-call tool exact (n = 39)	0.897	0.923	0.949
median seconds per answer	1.89	3.68
completion tokens per answer	219	377

The composite move is +0.014 [−0.006, +0.033]. That misses the pre-registered +0.05 bar. The gains that clear their intervals are CVE→CWE (+0.155), CVSS (+0.097), and the native tool-call loop (+0.310). Median answer time is about half. Token count is about half. Decode speed is the same (~110 tok/s).

On the same test items, against the other models that finished: wins over Cyber-Ornith-9B, Foundation-Sec-8B-Instruct, Primus-Merged-8B, SecGPT-7B, and the Qwen3.5-9B base. Ties with ab9b-r1 and with Huihui-Qwen3.8-27B-abliterated. The 27B tie is at roughly twice the speed per quality point. Full table: evidence/results/league_verdict.txt.

Training

One LoRA round continued from ab9b-r1. r = 32, α = 32, lr 1e-4 cosine, one epoch, 1,063 steps, 16.06M tokens, 4.2 hours on one RTX 3090 (24 GB). Data is defensive: NVD CVE classification, vulnerability review, threat-intel Q&A, ATT&CK / CAPEC / CWE, analyst-brief rewrites, and oracle-checked agent episodes. Nothing from the reserved eval families.

The shipped checkpoint is v2a-ck900. Two later rounds and two weight interpolations lost the pre-registered dev rule. v2b raised the agent scores and broke tool-call hygiene (it invents missing arguments). v2c repaired that hygiene and still scored below v2a-ck900, so it was never test-scored. Both adapters are in the repo so those tables can be checked.

Protocol: evidence/GOAL.md. Numbers: evidence/REPORT.md.

Files

ab9b-v2-Q4_K_M.gguf                 merged Q4_K_M, 5.8 GB
adapters/v2a-ck900/                 LoRA merged into the GGUF
adapters/v2b-final/                 later round, not selected
adapters/v2c-final/                 later round, not selected
evidence/GOAL.md                    protocol, written before the test pass
evidence/REPORT.md                  test table, league, rejected rounds
evidence/results/                   raw pick and verdict files
SHA256SUMS                          GGUF and the three adapter weights

Limits

Agent splits are small (89 / 87 / 39). Dev selection used 60 agent items, so a few points of noise is normal. Training families and the agent test share generators; eight reserved families were never trained on. CVE rows are the same task type as CTI-RCM and CTI-VSP, other CVEs, not the same items. Thinking mode, context past 64k, and languages other than English were not measured. The data is synthetic and defensive. It was not red-teamed.

License

Apache-2.0, same as Qwen3.5-9B. Public sources in the mix keep their own licences.

@software{ab9b_v2_2026,
  title  = {ab9b-v2},
  author = {dorb},
  year   = {2026},
  url    = {https://huggingface.co/dorb/ab9b-v2}
}