Qwen3-4B-Instruct-2507-Math-Fr

This model is obtained by merging SamsungSAILMontreal/Qwen3-4B-Instruct-2507-Math, SamsungSAILMontreal/Qwen3-4B-Instruct-2507-Code and SamsungSAILMontreal/Qwen3-4B-Instruct-2507-Fr. The model is used in the experiments described in https://bknyaz.github.io/blog/2026/meta-merge/. Single A100 was used for merging and evaluation.

The following versions were used for merge/eval:

python >= 3.10
torch : 2.9.0+cu128
lm_eval : 0.4.9.1
vllm : 0.11.1
transformers : 4.57.6
datasets : 3.2.0
numpy : 2.2.6

Merging

Merging was done using parameter averaging implemented in merge_qwen.py.

Evaluation

Evaluation was done with lm_eval on the test split of gsm8k, french_bench (avg score), gsm8k-fr and humaneval (instruct):

python -m lm_eval --model vllm --model_args pretrained=${model},tensor_parallel_size=1,dtype=auto,gpu_memory_utilization=0.9,data_parallel_size=1 \
 --tasks gsm8k,french_bench,gsm8k-fr,humaneval_instruct --batch_size 1 --apply_chat_template=True --confirm_run_unsafe_code --trust_remote_code

To evaluate on gsm8k-fr you can use our fork https://github.com/bknyaz/lm-evaluation-harness/tree/main/lm_eval/tasks/gsm8k.

Results

Model	gsm8k	french	gsm8k-fr	humaneval_instruct	avg
Qwen3-4B-Instruct-2507	80.4	43.1	66.0	90.2	69.9
Qwen3-4B-Instruct-2507-Math	76.8	43.0	65.3	72.0	64.3
Qwen3-4B-Instruct-2507-Fr	72.3	45.7	60.7	74.4	63.3
Qwen3-4B-Instruct-2507-Code	72.5	45.4	53.0	76.2	61.8
Qwen3-4B-Instruct-2507-Math-Code-Fr	82.9	45.8	69.8	79.9	69.6