xthi0p

gemma4-31b-it-amharic-finetome-lora-v1

README

License: apache-2.0

Acknowledgments and provenance

Base model: Google Gemma 4 31B Instruct, pinned to revision 842da3794eaa0b77d5f08bae87a17459d91ff475 (Apache 2.0).
Training dataset: AddisAI — FineTome single-turn deduplicated Amharic, pinned to revision fac04d882e094728d684275f296518a88e5cd4bd (dataset card lists Apache 2.0).
Dataset creation credit: AddisAI; Amharic examples are machine-translated using AddisAI's Aleph (፩) model from primarily English instruction sources.
The model was trained on a separately filtered and deduplicated subset, not the full AddisAI dataset. No original data or held-out questions are redistributed here.

Fine-tuning configuration

Table with columns: Setting, Value
Setting	Value
Method	BF16 LoRA, no 4/8-bit quantization
LoRA rank	16
Trainable LoRA parameters	122,429,440
Training examples (after length filter)	16,944
Validation examples (after length filter)	788
Max sequence length	1,024 tokens
Learning rate (peak)	2e-5
Gradient accumulation	8
Epochs	1
Optimizer steps	2,118
Final validation loss	0.8111
Final validation token accuracy	78.51%

These validation scores refer to the held-out portion of the locally filtered translated FineTome dataset. They do not establish general Amharic quality or superiority to the base model; independent evaluations have not yet been reported.

Limitations

Machine translation quality and Amharic linguistic naturalness have not been fully verified by native speakers.
String-similarity and numeric-template checks were applied to the local dataset; semantic and cross-language leakage were not fully ruled out.
Long examples over 1,024 tokens were excluded.
The adapter targets language-model layers; multimodal improvements are not claimed.
A capable GPU and the separately obtained Gemma 4 31B IT base model are needed to use it.

Loading

Load google/gemma-4-31B-it using the model's official instructions, then apply this adapter using Hugging Face PEFT's PeftModel.from_pretrained. Use a compatible Transformers/PEFT stack and the base model revision above.

License

Apache License 2.0. See LICENSE in this repository and the base model and dataset licenses. Respect the licenses and rights of original upstream data sources.

Adapter SHA-256: c84ebfae474260cc4b567ed191d4b1e19c3ec67f990c7e8ca5a9963d8277e7db.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

xthi0p

Model Tree

Base

google/gemma-4-31B-it

Adapter

this model

Input Modalities

TextImage

Output Modalities

Text

Supported Functionality