Chain-of-Thought Augmented Fine‑Tuning of a Distilled Llama‑8B Model for SIEM Detection Query Generation
Bibliographic record
Abstract
Writing effective security detection rules for SIEM systems is a complex and timeconsuming task that traditionally requires deep domain expertise.One of the persistent challenges in applying AI to security operations (SecOps) is the scarcity of high-quality, domain-specific datasets that can support the development of accurate and reliable models.This paper addresses that gap by presenting a method to both generate such a dataset and leverage it to fine-tune a compact reasoning-oriented model (DeepSeek R1 Distill Llama 8B) using Low Rank Adaptation (LoRA).As part of our contributions, we detail the creation of a curated, high-quality (of 1206 entries from 106 detection rule) dataset specifically tailored for SIEM text-to-query tasks, which enabled effective fine-tuning of the model.A key feature of this dataset is the augmentation of each training example with chain-ofthought (CoT) rationales: step-by-step explanations linking the natural language description of a detection rule to the resulting Lucene query.These rationales, produced by a stronger teacher model (Claude sonnet 4), are used to supervise the smaller student model.We describe the data pipeline, prompt templates, and LoRA configuration, and we report an initial human evaluation showing that CoT augmentation improves the reliability of textquery generation without increasing computational cost.Despite its compact size, our fine-tuned model outperformed several large proprietary language models in both query accuracy and reasoning quality.DeepSeek R1 Distill Llama 8B was chosen as a small model with reasoning capabilities, while Claude Sonnet 4 was selected for its strong ability to generate rationales.However, the research remains applicable to other small reasoning models and large, more capable models, respectively.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.001 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.000 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.001 | 0.000 |
| Scholarly communication | 0.000 | 0.003 |
| Open science | 0.001 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one teacher head, not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".