Description
Qwen2.5-Coder-7B-Instruct trained on a merged dataset of Unity3d q&a from these three datasets:
ibranze/codellama_unity3d_v2 (Full)
Hypersniper/unity_api_2022_3 (10%)
neph1/Unity_Code_QnA (Full)
preview 2: 26210 rows, of which ca 1000 are from my own multi response dataset
preview 1: 15062 rows in total with a 10% validation split.
Trained with native chat template (minus tools usage, see this issue: https://github.com/unslothai/unsloth/issues/1053). With a little superficial testing done, it seems to respond well to the mistral template.
Consider this a preview while I develop a dataset of my own.
If you have any feedback, please share. I've only done some basic testing so far. I'm especially interested if you're using it with Tabby or a similar coding tool.
Uploaded model
- Developed by: neph1
- License: apache-2.0
- Finetuned from model : unsloth/Qwen2.5-Coder-7B-Instruct-bnb-4bit
This qwen2 model was trained 2x faster with Unsloth and Huggingface's TRL library.
Training details
About 1.5 epochs. It's probably a bit overfitting and I should introduce some general coding questions to my validation set to ensure it doesn't lose too much general performance.
Rank: 128
Alpha: 256
TrainingArguments( per_device_train_batch_size =2, gradient_accumulation_steps = 64, #max_steps=10, num_train_epochs=3, warmup_steps = 5, learning_rate = 1e-4, fp16 = not torch.cuda.is_bf16_supported(), bf16 = torch.cuda.is_bf16_supported(), logging_steps = 10, optim = "adamw_8bit", weight_decay = 0.01, lr_scheduler_type = "linear", seed = 3407, per_device_eval_batch_size = 2, eval_strategy="steps", eval_accumulation_steps = 64, eval_steps = 10, eval_delay = 0, save_strategy="steps", save_steps=25, report_to="none", ),
Step Training Loss Validation Loss
20 2.043000 1.197104
40 1.087300 0.933553
60 0.942200 0.890801
80 0.865600 0.866198
100 0.851400 0.849733
120 0.812900 0.837039
140 0.812400 0.827064
160 0.817300 0.818410
180 0.802600 0.810163
200 0.788600 0.803399