Views
No views yet
google/gemma-4-E2B-it
for on-device function calling in the AI Business Card Holder Android app.find_contact(name, company?) — locate an existing contactadd_contact(name, phone?, email?, title?, company?) — create a new contactupdate_contact(id, fields) — edit an existing contactlist_contacts(filter?) — list / search contacts| Base model | google/gemma-4-E2B-it (2.3 B effective / 5.1 B total) |
| Method | LoRA via Unsloth |
| Dataset | ai-colombia/card-holder-finetune-data (8 000 multi-turn examples, 6 languages) — private |
| Rank / α / dropout | 16 / 16 / 0.05 |
| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
| Effective batch size | 16 (2 × 8 grad-accum) |
| Epochs / steps | 2 / 1 000 |
| Learning rate | 5e-5, linear, 20 warmup steps |
| Optimizer | adamw_8bit |
| Precision | fp16 (Colab T4) |
| Sequence length | 1024 |
| Trainable params | 31 M / 5.15 B (≈ 0.60 %) |
| Final training loss | 0.7541 |
| Train time | ≈ 113 min on Colab T4 |
| Date | April 2026 |
ℹ️ For mobile deployment, use the GGUF Q4_K_M quantization (see Files and versions once uploaded). The merged fp16 weights here are the source for that quantization.
1<bos><start_of_turn>user
2You are the Card Holder assistant. Use tools when needed.
3
4Find John Smith from Microsoft.<end_of_turn>
5<start_of_turn>model