NightBaRron1412/deepparse-r1-8b

DeepParse — DeepSeek-R1-Distill-Llama-8B LoRA Adapter

LoRA adapter for DeepParse, the hybrid log-parsing system from the EASE 2026 paper "DeepParse: Hybrid Log Parsing with LLM-Synthesized Regex Masks."

The adapter fine-tunes deepseek-ai/DeepSeek-R1-Distill-Llama-8B to emit a small JSON bundle of labelled regex masks when shown a sample of raw log lines. Those masks are consumed by the deterministic Drain stage at execution time — the LLM is invoked once at installation, never per log line.

Quick start

Via the DeepParse CLI

pip install deepparse
export HF_TOKEN=<your-token>   # base model is gated

deepparse synth \
    --mode hf \
    --adapter NightBaRron1412/deepparse-r1-8b \
    --input partner_logs.txt \
    --out masks.json

deepparse parse --masks masks.json --input partner_logs.txt --out parsed.csv

Via the Python API

from deepparse import synth_masks, Drain

with open("partner_logs.txt") as f:
    logs = f.read().splitlines()

patterns = synth_masks(
    logs,
    mode="hf",
    adapter_path="NightBaRron1412/deepparse-r1-8b",
)

drain = Drain()
drain.load_masks(patterns)
parsed = drain.parse_all(logs)

Hardware requirements

The adapter itself is ~80 MB, but it must be loaded on top of the 8 B base model:

Precision	VRAM (base)	Notes
bf16	~16 GB	Recommended; matches training recipe
fp32	~32 GB	Default in the DeepParse CLI; opt-in bf16 via --bf16
4-bit	~5 GB	Requires bitsandbytes; not wired into DeepParse CLI yet

The base model (deepseek-ai/DeepSeek-R1-Distill-Llama-8B) is gated on HuggingFace — request access on the model page and set HF_TOKEN before first use.

Training details

Hyperparameter	Value
Base model	deepseek-ai/DeepSeek-R1-Distill-Llama-8B
LoRA rank / alpha / dropout	8 / 32 / 0.01
Optimiser	AdamW, lr=2e-4 (cosine decay)
Batch size / grad accum	8 / 4
Epochs	25
Precision	bfloat16
Attention	eager (AMD ROCm cross-vendor stability)
Max sequence length	512
Training set	619 paper-curated examples
Hardware	AMD MI300A (ROCm 7.3)

Hyperparameters match the paper, Section "LLM Configuration and Fine-Tuning."

Intended use & limitations

In scope. Synthesising regex masks for the canonical variable classes covered by DeepParse (timestamps, IPs, MAC addresses, numerics, hex IDs, log levels, paths). The LLM is invoked once per log corpus at installation time; runtime parsing is fully deterministic.

Out of scope. Direct conversational use, code generation, anything outside the DeepParse pipeline. The model is fine-tuned to respond in a specific JSON schema and will not behave as a general-purpose assistant.

Limitations. Quality on log formats with unusual or domain-specific variable classes (custom IDs, vendor-specific tokens) may be lower than on the LogHub-style corpora the paper evaluates against. For those cases, consider DeepParse's offline stub mode (--mode offline) as a deterministic baseline or fine-tune further on a representative sample.

Reproducibility

• Source repo: https://github.com/NightBaRron1412/DeepParse
• Tagged release: v1.0.0
• Training script: python -m deepparse.training.finetune --train artifacts/training/train_paper.jsonl --output-dir artifacts/checkpoints/deepparse-r1-8b --epochs 25 --bf16

Citation

@inproceedings{deepparse2026,
  title  = {DeepParse: Hybrid Log Parsing with LLM-Synthesized Regex Masks},
  author = {Shetaia, Amir and Ratkov, Mihailo and Kauffman, Sean},
  booktitle = {Proceedings of EASE 2026},
  year   = {2026}
}

License

Apache 2.0, matching the DeepParse source repository. The base model retains its own licence (DeepSeek model licence); please review before redistribution.