cyber-byte-sft

Cybersecurity SFT mix for the byte-level model Bwen-8B-Cyber (fine-tune of allenai/Bwen-8B, Bolmo architecture). 33,050 instruction rows, attack-heavy, built with dedup + quality gates. data/tokens/ contains the pre-tokenized byte-level tensors (byte ids, subword-suffix expanded ids, true patch starts, assistant masks) ready for Bolmo/BLT-style training.

Mix

source	rows	share	role	license
oyildirim/cyberstrike-sft-120k	14,940	45.2%	red team / offense	Apache-2.0
Luoberta/CTF-Instruct (cyber-keyword filtered: only ~29% of the raw set is cyber)	2,447	7.4%	CTF	unspecified
self-built from DAPT-2020 pcaps (tshark)	3,303	10.0%	network forensics	—
Trendyol Cybersecurity v2.0 + Fenrir v2.1	7,920	24.0%	blue team / defense	Apache-2.0
reloading0101/threat-intelligence-dataset	4,440	13.4%	CTI	CC-BY-NC-SA-4.0

Build process

• Exact + near-duplicate dedup (first-user-message hash + prefix hash)
• Quality gates: length (200–14000 chars), min assistant content, unique-line-ratio ≥ 0.2 (pathological repetition), printable-ratio ≥ 0.85
• Literal-\n escape repair for Trendyol/Fenrir (double-escaped newlines, ≥3 occurrences)
• Cyber-keyword filter for CTF-Instruct (71% of rows are off-topic general coding questions)
• pcap portion generated with tshark from real captures: single-packet hex-dump decoding, conversation statistics, port-scan/beaconing anomaly detection, TCP-stream analysis; all answers computed from the actual parsed packets

Tokenized format (data/tokens/*.parquet)

Columns: byte_ids (int32, with BOS/EOS), expanded_ids (subword-suffix ids per byte, Qwen3 tokenizer), patch_start (bool, first byte of each subword patch), assistant_mask (bool). Chat template: User: {q}\n\nAssistant: {a} (+ optional system preamble). Compression: 3.86 bytes/patch.