---
license: apache-2.0
license_link: https://ai.google.dev/gemma/docs/gemma_4_license
library_name: gguf
pipeline_tag: translation
base_model:
- google/gemma-4-E4B-it
new_version: TRACCERR/Gemma-4-E4B-IT-Sumtablets-GGUF
language:
- su
- en
tags:
- gguf
- gemma4
- sumerian
- cuneiform
- transliteration
- translation
- unsloth
datasets:
- colesimmons/SumTablets
metrics:
- accuracy
- loss
model-index:
- name: Gemma-4-E4B-IT-Sumtablets
  results:
  - task:
      type: translation
      name: Sumerian Cuneiform Transliteration
    dataset:
      name: colesimmons/SumTablets
      type: colesimmons/SumTablets
      split: test
    metrics:
    - type: loss
      value: 0.30
      name: Train Loss (optimal stopping point)
    - type: character_accuracy
      value: 36.2
      name: Character-Level Accuracy
---

# Gemma-4-E4B-IT-Sumtablets-GGUF

**Gemma 4 E4B-IT** fine-tuned for **Sumerian cuneiform transliteration** — converting glyph name sequences into standard Latin-alphabet transliterations as used in Assyriology.

Trained on the [SumTablets](https://huggingface.co/datasets/colesimmons/SumTablets) dataset (82,452 administrative texts, predominantly Ur III period) using QLoRA via [Unsloth](https://github.com/unslothai/unsloth).

[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)

## Task

Given a sequence of Sumerian cuneiform **glyph names** (e.g., `1(diš) 1(aš) gur še lugal`), the model produces the corresponding **transliteration** in standard Assyriological convention.

This is a structured sequence-to-sequence translation task — not free-form text generation. The model must learn the compositional rules of Sumerian sign reading, including:

- Determinatives (e.g., `d`, `diš`)
- Logographic vs. syllabic sign readings
- Numeric notation systems (Š, N, etc.)
- Signs with multiple possible readings

## Training Details

| Parameter | Value |
|-----------|-------|
| **Base Model** | `unsloth/gemma-4-E4B-it` (4-bit QLoRA) |
| **Dataset** | `colesimmons/SumTablets` (82,452 train / 4,577 val / 4,577 test) |
| **Format Mapping** | `glyph_names → user`, `transliteration → assistant` |
| **System Prompt** | *"Transliterate the following Sumerian cuneiform glyph names into their corresponding Latin alphabet transliteration. Output ONLY the transliteration."* |
| **LoRA Rank** | 16 |
| **LoRA Alpha** | 16 (rsLoRA enabled) |
| **Learning Rate** | 5e-5 |
| **Scheduler** | Cosine |
| **Batch Size** | 2 × 8 accum = 16 effective |
| **Max Seq Length** | 4096 |
| **train_on_completions** | false |
| **packing** | false |
| **Hardware** | NVIDIA RTX 5080 (16GB) |

## Results

| Metric | Value |
|--------|-------|
| **Train Loss (optimal)** | 0.30 |
| **Character Accuracy** | 36.2% |

The 36% character-level accuracy reflects the inherent difficulty of the task — Sumerian sign readings are heavily context-dependent, with many signs having 5+ valid readings depending on genre, period, and surrounding signs. The model performs strongest on short administrative texts (Ur III, ~93% of training data) and weakest on long literary compositions.

**Note on inference:** Gemma-4-E4B's built-in thinking/reasoning mode can override the fine-tuned transliteration behavior. Use the same system prompt during inference as was used during training to suppress reasoning mode and get clean transliteration output.

## GGUF Quantizations

| File | Format | Size | Use Case |
|------|--------|------|----------|
| `gemma-4-e4b-it.Q4_K_M.gguf` | Q4_K_M | 4.97 GB | General use — best quality/size balance |
| `gemma-4-e4b-it.BF16-mmproj.gguf` | BF16 | 374 MB | Multimodal projector (required for vision input) |

## Usage

### llama.cpp / LM Studio

```bash
llama-cli \
  -m gemma-4-e4b-it.Q4_K_M.gguf \
  --system-prompt "Transliterate the following Sumerian cuneiform glyph names into their corresponding Latin alphabet transliteration. Output ONLY the transliteration." \
  -p "1(diš) gin₂ ku₃-babbar"
```

### llama-server (OpenAI-compatible)

```bash
llama-server \
  -m gemma-4-e4b-it.Q4_K_M.gguf \
  --system-prompt "Transliterate the following Sumerian cuneiform glyph names into their corresponding Latin alphabet transliteration. Output ONLY the transliteration."
```

## Limitations

- **Domain concentration**: 93% of training data is administrative texts from the Ur III period. Performance degrades on literary, lexical, or other period texts.
- **Character accuracy ceiling**: 36% char-level accuracy means roughly 1 in 3 characters will be incorrect — suitable as a pre-processing/assistive tool, not as a standalone authoritative transliteration.
- **Thinking mode interference**: Without the system prompt, the base model's reasoning mode may produce English analysis instead of transliteration.
- **No audio/vision evaluation**: The multimodal projector weights are included but were not evaluated for image-to-text cuneiform reading.

## Dataset

Training used [colesimmons/SumTablets](https://huggingface.co/datasets/colesimmons/SumTablets) (CC-BY-4.0):

- **82,452** training samples, **4,577** validation, **4,577** test
- Columns: `glyph_names` (input) → `transliteration` (target)
- Predominantly Ur III period administrative records
- Covers genres: administrative, legal, letter, literary, lexical, royal inscription, school text

## Acknowledgments

- **Google DeepMind** — Gemma 4 model family
- **Colesimmons** — SumTablets dataset
- **Unsloth** — QLoRA training framework (2× faster fine-tuning)

## License

Apache 2.0 — see [Gemma 4 License](https://ai.google.dev/gemma/docs/gemma_4_license) for model terms. Dataset licensed separately under CC-BY-4.0.