---
license: apache-2.0
datasets:
- SubMaroon/DTF_Comments_Responses_Counts
language:
- ru
base_model:
- unsloth/Qwen2.5-7B
pipeline_tag: text-generation
---

A continued pretrained version of unsloth/Qwen2.5-7B model using unsloth's low rank adaptation on a dataset of [DTF](dtf.ru) posts. The adapter is already merged with the model.

For pretraining, posts from [SubMaroon/DTF_comments_Responses_Counts](https://huggingface.co/datasets/SubMaroon/DTF_Comments_Responses_Counts) were selected, deduplicated by simple `df.unique` and filtered by length of 1000 < x < 128000 tokens. The training dataset size was roughly 75M tokens.

LoRA hyperparameters:

```
r=32
target_modules=[
    "q_proj", 
    "k_proj", 
    "v_proj", 
    "o_proj",
    "gate_proj",
    "up_proj",
    "down_proj",
]
lora_alpha=16
lora_dropout=0
bias="none"
use_gradient_checkpointing='unsloth'
use_rslora=True
random_state=42

```

Training hyperparameters:

```
num_train_epochs=2
train_batch_size=8
gradient_accumulation_steps=16
gradient_checkpointing=False
optim="adamw_8bit"
weight_decay=4e-2
bf16=True
learning_rate=5e-5
lr_scheduler_type="cosine"
packing=True,
seed=42
```

Training time:

- NVidia Tesla A100 80GB: ~8.5 hours
- NVidia RTX 3090ti: ~33.5 hours

[Wandb](https://wandb.ai/a_okshus/DTF_comments/runs/fr5hfq6g?nw=nwusera_okshus)

[GitHub: TODO]()