nesilabs/tosilos-24b-2512

Tosilos-24b-v2 — European cybersecurity model (Devstral-Small-2 24B 2512, fine-tuned)

Second-generation Tosilos 24B: QLoRA (r=64 alpha=128, 2 epochs) on Devstral-Small-2-24B-Instruct-2512 (dequantized FP8→bf16), trained on 12,201 examples (domain + 25% general replay, verified nuclei tasks included).

Benchmarks (same harness, 500 questions each)

Metric	Devstral-2512 base	Tosilos-24b-2512	Tosilos-24b (2505, reference)
CyberMetric	91.8%	91.6%	92.8%
MMLU (general)	73.0%	72.2%	77.4%
Domain judge (Opus 5, 1-10)	5.33	6.11 (+0.78, 28W/21L/6T)	5.68

Honest read: v2 ties its base on MCQ but shows the strongest domain-depth gain we have measured (+0.78 judge delta — richer dataset with verified nuclei tasks and the full AD corpus). Choose nesilabs/tosilos-24b (2505) for max general+knowledge balance, or this 2512 for max domain depth. Both keep general capability intact (MMLU within noise).

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "nesilabs/tosilos-24b-2512", load_in_4bit=True, device_map="auto")
tok = AutoTokenizer.from_pretrained("nesilabs/tosilos-24b-2512")

Training recipe

QLoRA r=64 alpha=128, targets q/k/v/o + gate/up/down, 4-bit NF4, seq 4096, 2 epochs on 12,201 examples, lr 1e-4 cosine, 1×H200, 1,411 steps. Base dequantized from the official FP8 release (verified fp8×scale → bf16).

Dual-use: intended for authorized security work only.