edorazio/gemma-3-270m-it-dynamicplp

Gemma 3 270M — DynamicPLP Profiler

Fine-tuned Gemma 3 270M IT for in-browser e-commerce user profiling. Designed to run client-side in DynamicPLP via Transformers.js + WebGPU.

Task

Analyzes shoe-browsing interaction logs and assigns preference weights across 3 dimensions:

• Colors (23 values): rosso, blu, nero, beige, ...
• Styles (6 values): casual, classic, elegant, minimal, sporty, urban
• Categories (6 values): flat, high_heel, hiking_boot, mans_shoe, running, womans_boot

Output format (6 lines, Italian):

COLOR colore=peso, colore=peso
STYLE stile=peso, stile=peso
CATEGORY categoria=peso, categoria=peso
CONFIDENCE 0.0-1.0
INTENT exploring|deciding|focused
MESSAGE <1-2 Italian sentences>

Training

• Teacher: mlx-community/gemma-4-e4b-it-4bit (Gemma 4 E4B with thinking enabled)
• Dataset: 1071 synthetic browsing sessions, strictly whitelist-filtered
• Method: LoRA (rank=16, scale=2.0) on q/k/v/o projections, all layers
• Optimizer: AdamW, lr=1e-4, 1000 iters, batch=4
• Framework: mlx-lm → ONNX via optimum

Eval (161 holdout examples)

Metric	Value
Format compliance	100%
Invalid labels	2.5%
Color cosine sim vs teacher	0.72
Style cosine sim vs teacher	0.77
Category cosine sim vs teacher	0.78
Intent accuracy (deciding)	100%
Confidence MAE	0.061
Gen speed (MLX bf16)	0.58s/ex

Known limitation: training data was 91% deciding intent. Student collapses minority intents (exploring, focused → 0% recall). Reordering quality (weights) unaffected; intent only impacts trigger cooldown in DynamicPLP.

Files

File	Size	Use
onnx/model.onnx	1.6 GB	fp32 reference
onnx/model_fp16.onnx	1.3 GB	CPU fallback
onnx/model_q4.onnx	800 MB	fp32 activations + 4-bit weights
onnx/model_q4f16.onnx	481 MB	WebGPU primary

Usage in Transformers.js

import { pipeline } from '@huggingface/transformers';

const pipe = await pipeline(
  'text-generation',
  'edorazio/gemma-3-270m-it-dynamicplp',
  { dtype: 'q4f16', device: 'webgpu' }
);

License

Gemma Terms of Use. See https://ai.google.dev/gemma/terms