gemma-4-E2B-sft-tran-tigre

GGUF builds of a Tigre translation fine-tune of google/gemma-4-E2B.

Pipeline

1. Base — google/gemma-4-E2B (Gemma 4 E2B)
2. CPT — Continued pre-training adapter for Tigre fluency, merged into base weights
3. SFT — Translation LoRA on bidirectional translation pairs (X↔Tigre), merged into the CPT-merged base
4. GGUF — Converted with llama.cpp (convert_hf_to_gguf.py) and quantized with llama-quantize

Files

File	Quantization	Notes
gemma-4-E2B-sft-tran-tigre-BF16.gguf	BF16	Full precision base — use to derive other quants
gemma-4-E2B-sft-tran-tigre-Q8_0.gguf	Q8_0	Near-lossless, recommended if size allows
gemma-4-E2B-sft-tran-tigre-Q5_K_M.gguf	Q5_K_M	Good quality, smaller than Q8
gemma-4-E2B-sft-tran-tigre-Q4_K_M.gguf	Q4_K_M	Best size/quality tradeoff for most users

Usage with llama.cpp

./llama-cli -hf google/gemma-4-E2B-sft-tran-tigre:Q4_K_M \
    --temp 0.3 --top-p 0.9 \
    -p "Translate the following English text to Tigre.\n\nHello, how are you?"

Chat template

Standard Gemma turn template:

<bos><start_of_turn>user
{instruction}

{input}<end_of_turn>
<start_of_turn>model
{output}<end_of_turn>

License

Inherits the Gemma Terms of Use from the base model.