Qwen3.6-35B-A3B-Tool-Slo-LoRA-v2

Qwen3.6-35B-A3B Tool-Slo medical — experimental LoRA v2

Trained by MediaAtlas — LLM fine-tuning on your own data, trained in the EU, weights delivered. Pricing · All our models

Na kratko: Poskusni adapter LoRA za Qwen/Qwen3.6-35B-A3B: kratek dodatni SFT na sintetičnih slovenskih medicinskih robnih primerih z naborom za ohranjanje pogovora. Ni evalviran; za rabo priporočamo texdata/Qwen3.6-35B-A3B-Tool-Slo-GGUF.

This model is a fine-tuned version of Qwen/Qwen3.6-35B-A3B on the sft_synthetic_med_edge and the sft_chat_ret datasets. It achieves the following results on the evaluation set:

• Loss: 1.0797

Model description

An experimental LoRA adapter from a short supervised fine-tuning run (24 optimizer steps over 3 epochs, effective batch 32) on the internal sft_synthetic_med_edge set (synthetic Slovenian medical examples) plus sft_chat_ret (Slovenian chat retention). Published for transparency and reproducibility of our training runs.

Intended uses & limitations

Research only. Not evaluated beyond validation loss, not a medical device, not for clinical use; outputs may be wrong. For production-quality models see the main cards linked above.

Training and evaluation data

Internal sets sft_synthetic_med_edge and sft_chat_ret; held-out validation split of the same data (loss 1.0797). No external benchmark was run.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• gradient_accumulation_steps: 8
• total_train_batch_size: 32
• total_eval_batch_size: 4
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 0.05
• num_epochs: 3.0

Training results

Framework versions

• PEFT 0.18.1
• Transformers 5.6.0
• Pytorch 2.12.1+cu130
• Datasets 4.0.0
• Tokenizers 0.22.2