GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Sinall
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Sinall
Model Tree
google/gemma-4-12B
Input Modalities
Output Modalities
Supported Functionality
WIKIPEDIA-SV), sampled through Språkbanken's Korp API. Språkbanken lists the corpus as CC-BY-4.0.842da3794eaa0b77d5f08bae87a17459d91ff475), run through NeMo DataDesigner. The method adapts the source-grounded generation and curation of Source2Synth. A candidate was kept only if an independent answer matched it exactly after normalisation and it passed a quality review. Duplicates were then removed.| Task | Rows |
|---|---|
| classification | 143 |
| sentiment classification | 143 |
| linguistic acceptability | 143 |
| reading comprehension | 143 |
| natural language inference | 143 |
| grammatical error detection | 142 |
| named-entity recognition | 104 |
The named-entity task stopped short of its quota because most of its later candidates duplicated earlier ones. The training data itself is not published here.
| Setting | Value |
|---|---|
| Method | bf16 LoRA (no quantisation) |
| LoRA rank / alpha / dropout | 16 / 32 / 0.05 |
| Target modules | q, k, v, o, gate, up, down projections |
| Learning rate, warmup | 2e-4, 5% |
| Effective batch size | 16 (8 × 2 accumulation) |
| Epochs, steps | 1, 61 |
| Max length | 2048 tokens |
| Loss | assistant answer and EOS only; prompt tokens masked |
| Final training loss |
Base Gemma 4 12B has no chat template, so training used the plain role template in chat_template.jinja (USER: / ASSISTANT: turns). Use the same template when prompting.
These are pilot numbers: EuroEval few-shot as a base model on the validation split with one iteration, and the full Swedish idioms minimal-pair set. They are not comparable to the EuroEval leaderboard.
| Benchmark | Base | Base + this adapter |
|---|---|---|
EuroEval swerec (MCC) | 77.3 | 74.1 |
EuroEval scala-sv (MCC) | 78.0 | 69.4 |
| Idioms, all 478 pairs (accuracy) | 91.2% | 90.2% |
| Idioms, 100-pair challenge set (accuracy) | 87% | 85% |
The idioms differences are not significant (exact McNemar p = 0.18 and 0.63). The very low final training loss suggests the adapter fits the synthetic answer format closely, which may not transfer to EuroEval's prompts.
python
from peft import PeftModelfrom transformers import AutoModelForImageTextToText, AutoTokenizer base = AutoModelForImageTextToText.from_pretrained("google/gemma-4-12B", dtype="bfloat16", device_map="auto")model = PeftModel.from_pretrained(base, "Sinall/gemma-4-12B-sv-source2synth-nlu-lora")tokenizer = AutoTokenizer.from_pretrained("Sinall/gemma-4-12B-sv-source2synth-nlu-lora")The adapter is released under Apache-2.0, the license of the base model. Training data was derived from Swedish Wikipedia (CC-BY-4.0 as listed by Språkbanken).
| 0.080 |
| Hardware | 1 × NVIDIA RTX PRO 6000 Blackwell |