divahno/cx-SmolLM2-360M-Q8_0

CX SmolLM2-360M Q8_0 GGUF

Fine-tuned SmolLM2-360M-Instruct for CX (Customer Experience) analytics insights — part of the Action-XM AI Guide system.

Model Details

Property	Value
Base model	HuggingFaceTB/SmolLM2-360M-Instruct
Architecture	LlamaForCausalLM
Parameters	360M
Quantization	Q8_0 (GGUF)
File size	369 MB
Context length	8192 tokens
Training framework	MLX-LM (Apple Silicon)

Training

• Method: LoRA (r=16, alpha=32, targets: q/k/v/o projections)
• Dataset: 9,828 synthetic CX analytics examples (ChatML format)
• Iterations: 1,000
• Learning rate: 2e-5
• Batch size: 2
• LoRA layers: 16 (of 32)
• Peak memory: 3.2 GB
• Hardware: Apple Silicon (MLX)

Training Data

Synthetic CX insight pairs generated via Claude Sonnet 4.6, covering:

• Funnel analysis and drop-off diagnosis
• Rage click / dead click interpretation
• Session replay pattern analysis
• Core Web Vitals optimization
• Scroll depth and engagement insights
• Heatmap and click pattern analysis
• Quick-back / bounce diagnosis
• Segment comparison and cohort analysis

Quality-gated: each example passed JSON structure, length, hallucination, and actionability checks.

Usage

llama.cpp

llama-cli -m cx-SmolLM2-360M-Q8_0.gguf -n 256 --temp 0.7 --chat-template chatml

Python (llama-cpp-python)

from llama_cpp import Llama

llm = Llama(model_path="cx-SmolLM2-360M-Q8_0.gguf", n_ctx=2048)
response = llm.create_chat_completion(messages=[
    {"role": "system", "content": "You are a CX analytics assistant."},
    {"role": "user", "content": "Cart abandonment is 71%. Average payment page duration is 12s. Insights?"},
])
print(response["choices"][0]["message"]["content"])

Performance

On Apple Silicon (M1 Pro):

• Prompt processing: ~579 tokens/sec
• Generation: ~164 tokens/sec

License

Apache 2.0 (same as base model)