KIEFERSA
Qwen3.5-0.8B-Greek
Greek fine-tuned Qwen3.5-0.8B — PyCon workshop

Qwen3.5-0.8B-Greek was made for the PyCon workshop. It is Qwen/Qwen3.5-0.8B fine-tuned for natural Greek chat and long conversations. Because of its small size, it has some limits in longer conversations (it can lose track of details or repeat itself); this is a limit of the model size.

It keeps the vision capabilities of Qwen3.5-0.8B: it also understands images and video.

MTP (self-drafting)

The model keeps its Multi-Token Prediction (MTP) layer, so it can draft its own next tokens for faster generation (speculative decoding, no extra draft model). The MTP layer was retrained on our chat data so its drafts match the fine-tuned model:

vllm serve KIEFERSA/Qwen3.5-0.8B-Greek --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":2}'

vLLM, 40 Greek prompts, greedy decoding, 2 draft tokens per step.

MTPOriginal MTP layerRetrained MTP layer
Acceptance rate32.3%58.4%
1st draft token accepted47.1%79.0%
2nd draft token accepted17.5%37.9%
Tokens per step1.652.17

Evaluation

Greek benchmarks, 0-shot, accuracy (%). Qwen3.5-0.8B-Greek highlighted.

BenchmarkQwen3.5-0.8B (base)Qwen3.5-0.8B-Greek
Average (all 6)34.8340.23
HellaSwag32.0433.58
Medical MCQA19.6828.94
ARC-Challenge34.2544.35
Belebele48.2252.78
TruthfulQA24.1131.21

Forgetting check

Benchmarks where the fine-tuned model scores slightly below the base model, kept here to track forgetting.

BenchmarkQwen3.5-0.8B (base)Qwen3.5-0.8B-Greek
Winogrande50.6750.51

GreekMMLU

GreekMMLUQwen3.5-0.8B (base)Qwen3.5-0.8B-Greek
Overall40.9846.55
Humanities41.3546.05
Art professional48.3352.00
Art secondary school30.7720.51
Art university58.3358.33
Greek history primary school36.6443.32
Greek history professional25.6431.62
Greek history secondary school38.7145.16
Greek literature35.7142.86
Greek mythology37.3944.96
Greek traditions45.4848.14
Prehistory55.5663.49
World history35.0045.00
World religions34.8438.06
Social sciences40.5347.13
Accounting35.3338.59
Economics professional30.6142.86
Economics university54.6565.12
Education professional52.1751.78
Education university31.7124.39
Geography primary school39.4451.39
Geography secondary school37.1033.87
Government and politics primary school57.1476.07
Government and politics secondary school38.6744.00
Law28.6334.51
Management professional51.7954.76
Management university52.0056.00
Modern greek language primary school35.8640.60
Modern greek language secondary school47.0556.70
Stem40.5646.53
Agriculture professional43.9549.26
Agriculture university38.5742.86
Biology28.4738.04
Chemistry33.3339.51
Civil engineering46.8149.28
Clinical knowledge44.0847.87
Computer networks and security33.7233.72
Computer science professional59.8471.65
Computer science university34.6242.31
Electrical engineering45.1146.01
Mathematics31.7537.57
Medicine professional42.4243.51
Medicine university34.7241.67
Physics primary school39.3465.57
Physics professional41.9246.88
Physics university33.8033.80
Other43.0745.62
Driving rules44.1545.72
General knowledge36.4745.58
Qa maritime safety and rescue operations46.6244.59

License

Inherits the Qwen3.5-0.8B base-model license (Apache 2.0).

Downloads last month
19
Safetensors
Model size
0.9B params
Tensor type
F32
·
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for KIEFERSA/Qwen3.5-0.8B-Greek

Finetuned
(857)
this model
Quantizations
1 model

Dataset used to train KIEFERSA/Qwen3.5-0.8B-Greek

Collection including KIEFERSA/Qwen3.5-0.8B-Greek