Tiny Aya SMS extraction: models
Collection
Lexsi hackathon T1-P5: LoRA + GGUF models β’ 10 items β’ Updated
How to use Meowniece/tiny-aya-sms-extract-v2-lora with PEFT:
from peft import PeftModel
from transformers import AutoModelForCausalLM
base_model = AutoModelForCausalLM.from_pretrained("CohereLabs/tiny-aya-global")
model = PeftModel.from_pretrained(base_model, "Meowniece/tiny-aya-sms-extract-v2-lora")LoRA adapter for CohereLabs/tiny-aya-global trained with AlignTune on data curated with CuratorKIT and evaluated with AuditKIT.
Input: one SMS in English, Hindi or Hinglish. Output: one JSON object with 11 fields (see the dataset card).
| Split | Exact match | Field accuracy | Hallucination β | Non-txn F1 |
|---|---|---|---|---|
| test_id | 92.0 | 98.7 | 0.2 | 100.0 |
| test_ood | 85.0 | 97.8 | 2.0 | 98.2 |
| challenge | 70.5 | 92.0 | 10.1 | 81.5 |
Training path: aligntune Β· prompt format: chat Β· 1600 curated examples Β· LoRA r=16, Ξ±=32, lr=0.0002, 3 epoch(s).
Licence: the base model is CC-BY-NC (non-commercial). Commercial deployment needs a licence from Cohere. Data: fully synthetic; bank names are fictional; no real customer data.