A fine-tuned language model for analyzing performance anomalies in distributed crypto exchange systems. Built for the KrystalineX observability platform.

Model Details

Property	Value
Base Model	meta-llama/Llama-3.2-1B-Instruct
Method	LoRA (Low-Rank Adaptation)
Trainable Parameters	1.56M / 1.24B (0.13%)
Training Framework	Axolotl
Precision	BF16 with 8-bit quantized base
License	Apache 2.0

Intended Use

This model analyzes OpenTelemetry trace data and correlated system metrics to identify root causes of performance anomalies in microservice architectures. Given an anomaly report containing span attributes, latency deviations, and system metrics (CPU, memory, error rates), the model produces:

Summary of the likely cause
Root causes with reasoning based on actual metric values
Actionable recommendations for remediation
Confidence level assessment

Example Input

Analyze this performance anomaly:
- Service: kx-exchange
- Operation: pg-pool.connect
- Duration: 286.94ms (expected: 0.44ms ± 10.66ms)
- Deviation: 26.88σ
- CPU Usage: 0.5%, Memory: 142MB, Error Rate: 0.0%

Example Output

SUMMARY: The pg-pool.connect operation experienced extreme latency due to
connection pool exhaustion requiring a new TCP connection to PostgreSQL.

CAUSES:
- Connection pool was empty, forcing a new connection establishment
- TCP connect span of 264ms confirms network-level connection setup
- Idle timeout (30s) likely evicted pooled connections

RECOMMENDATIONS:
- Increase minimum pool size to maintain warm connections
- Reduce idle timeout or implement connection keepalive
- Add connection pool metrics to monitoring

CONFIDENCE: high

Training Details

Dataset

222 training examples (22 real + 200 synthetic) of anomaly analysis from a production crypto exchange platform:

22 expert-curated examples from real OpenTelemetry traces, including hallucination corrections (e.g., teaching the model NOT to cite "high CPU usage" when CPU is at 0.5%)
200 synthetic examples generated across 15 anomaly scenario templates (connection pool exhaustion, cold start, query lock contention, DNS cache miss, network jitter, GC pause, message queue backlog, cascading timeout, retry storm, etc.)
~40% dismissal training: examples explicitly teaching the model to dismiss low metrics as irrelevant rather than hallucinating problems
Mixed prompt formats: both short-form (Analyze anomaly: service:operation) and detailed structured prompts with full span attributes and metrics

LoRA Configuration

Parameter	Value
Rank (r)	16
Alpha	32
Dropout	0.05
Target Modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj

Training Hyperparameters

Parameter	Value
Epochs	3
Learning Rate	2e-4
Scheduler	Cosine
Warmup Ratio	0.1
Batch Size	2 (micro) × 4 (grad accum) = 8 effective
Optimizer	AdamW
Sequence Length	2048

Training Results (v2 — 222 examples)

Epoch	Loss	Learning Rate
0.36	2.5158	1.83e-4
1.07	1.6355	1.00e-4
1.79	0.5071	2.66e-5
2.50	0.3211	4.88e-6
3.0	0.2422	0

Final Training Loss: 0.24 (down from 2.41 on 22 examples)
Training Time: ~56 minutes (84 steps) on NVIDIA Turing GPU (sm_75)
VRAM Usage: ~1.9GB training, ~4.4GB cache
Throughput: 0.2 samples/s, 0.025 steps/s

Usage

With Ollama

ollama run anomaly-analyzer "Analyze: service latency 500ms, expected 50ms, CPU 0.1%"

With Transformers

python

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("XavierThibaudon/anomaly-analyzer")
tokenizer = AutoTokenizer.from_pretrained("XavierThibaudon/anomaly-analyzer")

prompt = "Analyze anomaly: kx-exchange GET 500ms, expected 50ms, CPU 0.1%"
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Limitations

Trained on 222 examples (22 real + 200 synthetic) — results continue to improve with more real-world data
Optimized for the KrystalineX platform's specific service topology (kx-exchange, kx-wallet, api-gateway, order-matcher)
Best results when prompts include correlated system metrics alongside trace data
Small 1B model may not always follow strict output formatting — the parser handles free-form responses gracefully
May hallucinate metric interpretations for scenarios not represented in training data

Citation

bibtex

@misc{krystalinex-anomaly-analyzer,
  title={KrystalineX Anomaly Analyzer},
  author={Xavier Thibaudon},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/XavierThibaudon/anomaly-analyzer}
}