beau-warren/blueteam-cyber-base-v0.3.0

blueteam-cyber-base-v0.3.0

A roughly 100.7M-parameter Llama-compatible base model pretrained from scratch: a mixture of general English, code, cybersecurity text, defensive code-review examples, and Open Interpreter-style traces.

This is a base completion model, not an instruction-tuned assistant. Evaluate it as a text continuation model before applying supervised or preference fine-tuning.

Fine-tuning status: none. This repository contains the raw pretrained base checkpoint. It has not received supervised instruction fine-tuning, preference optimization, vulnerability-report fine-tuning, tool-call fine-tuning, or Open Interpreter integration. It should not be presented as a reliable security reviewer or autonomous agent.

Training summary

• —Parameters: 100,682,496
• —Training step: 57,308 / 57,308
• —Tokens processed: 1,877,868,544
• —Context used for pre-training: 1,024
• —Prepared training split: 1,877,928,931 custom-tokenizer tokens
• —Tokens packable at the configured sequence length: 1,877,897,216
• —Checkpoint slot: final

Base-model evaluation

Native evaluation completed on August 17, 2026:

• —Full prepared holdout: 20,008 sequences / 20,488,192 target tokens
• —Holdout coverage: 100% of packable targets
• —Cross-entropy loss: 1.9503
• —Perplexity: 7.0310
• —Qualitative result: partially coherent security vocabulary, but substantial repetition, factual errors, and broken code generation
• —Two-case structured security format check: 0% valid JSON and 0% verdict accuracy

The security check was only one vulnerable/patched SQL-injection pair and is a pipeline demonstration, not a statistically meaningful benchmark. The result confirms that this base checkpoint requires supervised instruction and report-format fine-tuning before it can be evaluated as a code-review assistant or tool-using agent.

GGUF and Ollama

blueteam-cyber-base-v0.3.0-Q8_0.gguf is the portable llama.cpp/Ollama build:

• —Format: GGUF V3 / Q8_0
• —Size: 107,838,208 bytes (102.07 MiB)
• —SHA-256: 4c6b00c9c4a84e04691f904638f68f5aaf00d418b0dff4c783952e22dccca45f
• —Converter: official llama.cpp revision d8df12ebc4592b55dcecf97a32953623d031efdc
• —Validation: loaded by Ollama 0.13.5, all 13 layers offloaded to the RTX 3060, and a bounded generation smoke test completed successfully

Download the GGUF and Modelfile into the same directory, then run:

ollama create blueteam-cyber-base-v0.3.0 -f Modelfile
ollama run blueteam-cyber-base-v0.3.0

This only changes the runtime format. Quantization does not add instruction-following, chat, vulnerability-review, or tool-use capabilities.

This corpus contains security-related code and strings that can trigger malware scanners, but the training pipeline treats JSONL content as inert text. The security/agent/report portions are predominantly synthetic. Test factuality, memorization, security behavior, and code quality before downstream use.