---
license: apache-2.0
base_model:
  - google/gemma-4-E2B-it
base_model_relation: finetune
datasets:
  - alwaysgood/financial-english-source-corpus
  - alwaysgood/financial-english-source-corpus-gemma4-e2b-1280
language:
  - en
  - ko
library_name: transformers
pipeline_tag: any-to-any
tags:
  - translation
  - financial-translation
  - english-to-korean
  - conversational
  - multimodal
  - safetensors
  - gguf
  - llama-cpp
  - lm-studio
  - q4_k_m
---

# Gemma4_E2B_ADS

Gemma4_E2B_ADS is a full fine-tune of
[`google/gemma-4-E2B-it`](https://huggingface.co/google/gemma-4-E2B-it) for
English-to-Korean financial translation. It was trained with the DQS
low-QE curriculum using seed 42.

This repository contains both the original Transformers checkpoint and one
LM Studio/llama.cpp export:

- `model.safetensors`: original BF16 fine-tuned checkpoint
- `Gemma4_E2B_ADS-Q4_K_M.gguf`: the only quantized main-model variant
- `mmproj-Gemma4_E2B_ADS-BF16.gguf`: multimodal encoder/projector companion

The BF16 `mmproj` is not an additional LLM quantization variant. It is kept at
BF16 for multimodal compatibility and quality.

## LM Studio

Use the latest LM Studio runtime and download the Q4_K_M variant:

```bash
lms get https://huggingface.co/alwaysgood/Gemma4_E2B_ADS@Q4_K_M
```

The matching `mmproj` file enables image input. Direct llama.cpp multimodal
testing with this checkpoint requires `--jinja`. Gemma 4 audio support may vary
by runtime and is not guaranteed by this model card. For translation, disable
thinking and ask for translation-only output, for example:

```text
Translate the following English financial text into Korean. Return only the translation.

<source text>
```

## Training and provenance

- Tuning: full-parameter supervised fine-tuning
- Seed: 42
- Selection: low quality-estimation score first (`qe_selection_order=low`)
- Base model thinking during training/evaluation: disabled
- Vision/audio layers: not trained; the base model's multimodal components were preserved
- Run artifacts: [`gemma4_e2b_it_full_lowqe_seed42`](https://huggingface.co/datasets/alwaysgood/dqs-runs/tree/fa8166a883d96460cc285b46d66b74a074b4b8d4/gemma4_e2b_it_full_lowqe_seed42)
- Source revision: `fa8166a883d96460cc285b46d66b74a074b4b8d4`

## Evaluation

The following scores are from the original BF16 final checkpoint on the
500-row held-out test set. They are not claimed as a separate Q4_K_M evaluation.

| Metric | Score |
|---|---:|
| BLEU | 30.7621 |
| chrF | 49.3295 |
| COMET (`wmt22-comet-da`) | 0.8968 |
| COMETKiwi (`wmt22-cometkiwi-da`) | 0.8630 |
| XCOMET-XXL | 0.8746 |
| MetricX-24 Hybrid XXL (lower is better) | 3.4078 |

Full evaluation records and configuration are available in the linked run.

## License and data note

The model weights follow the Apache-2.0 license of the base model. The training
corpus aggregates sources with mixed upstream terms; the dataset card is marked
`license: other`. Users are responsible for reviewing the source-specific terms
described in [`alwaysgood/financial-english-source-corpus`](https://huggingface.co/datasets/alwaysgood/financial-english-source-corpus).