README.md

---
library_name: peft
license: apache-2.0
base_model: Qwen/Qwen2.5-Coder-7B-Instruct
tags:
- axolotl
- base_model:adapter:Qwen/Qwen2.5-Coder-7B-Instruct
- lora
- transformers
- roblox
- luau
datasets:
- darwinkernelpanic/luau_corpus_axolotl
pipeline_tag: text-generation
model-index:
- name: Qwen2.5-Coder-7B-Instruct-Luau
  results: []
---

[<img src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/axolotl-ai-cloud/axolotl)
<details><summary>See axolotl config</summary>

axolotl version: `0.13.0.dev0`
```yaml
base_model: Qwen/Qwen2.5-Coder-7B-Instruct
# Auto-upload to HuggingFace when done
hub_model_id: darwinkernelpanic/Qwen2.5-Coder-7B-Instruct-Luau  # Change this to your HF username
hub_strategy: every_save  # Uploads checkpoints as you train
trust_remote_code: true

load_in_8bit: false
load_in_4bit: true

datasets:
  - path: darwinkernelpanic/luau_corpus_axolotl
    type: completion
    field_instruction: text  # Check the actual column names on HF
    field_output: completion   # Might be "text" or "code" — verify first

dataset_prepared_path:
val_set_size: 0.05
output_dir: ./outputs/qwen-luau-finetune

sequence_len: 2048
sample_packing: true
eval_sample_packing: true

adapter: qlora
lora_model_dir:
lora_r: 64
lora_alpha: 64
lora_dropout: 0.05
lora_target_linear: true

# Weights & Biases tracking (optional but clutch)
wandb_project: qwen-luau-finetune
wandb_entity:
wandb_watch:
wandb_name: qwen2.5-coder-7b-luau
wandb_log_model:

gradient_accumulation_steps: 2
micro_batch_size: 2
num_epochs: 3
optimizer: adamw_torch_fused
lr_scheduler: cosine
learning_rate: 0.0003
bf16: auto
tf32: true

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false

resume_from_checkpoint:
logging_steps: 10
flash_attention: true
warmup_ratio: 0.1
evals_per_epoch: 4
saves_per_epoch: 1
weight_decay: 0.01

fsdp:
  - full_shard
  - auto_wrap
fsdp_config:
  fsdp_limit_all_gathers: true
  fsdp_sync_module_states: true
  fsdp_offload_params: false
  fsdp_use_orig_params: false
  fsdp_cpu_ram_efficient_loading: true
  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
  fsdp_transformer_layer_cls_to_wrap: Qwen2DecoderLayer
  fsdp_sharding_strategy: FULL_SHARD
  fsdp_state_dict_type: FULL_STATE_DICT

special_tokens:
  pad_token: "<|endoftext|>"
```

</details><br>

# Qwen2.5-Coder-7B-Instruct-Luau

This model is a fine-tuned version of [Qwen/Qwen2.5-Coder-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct) on the darwinkernelpanic/luau_corpus_axolotl dataset.
It achieves the following results on the evaluation set:
- Loss: nan
- Ppl: nan
- Memory/max Active (gib): 14.12
- Memory/max Allocated (gib): 14.01
- Memory/device Reserved (gib): 14.69

## Model description

The model was fine-tuned on the Roblox/luau_corpus dataset which was converted to have the "prompt" collum replaced by "text" for compatibility reasons.
It was fine-tuned for improved knowledge and performance on Luau code (Roblox's Lua dialect, see [luau.org](https://luau.org)), which should end up improving code quality for Luau and Roblox projects.

## Intended uses & limitations

This model is intended for use within applications that use the Luau programming language, including but not limited to
- Roblox projects
- Standalone Luau projects (Lune?)
  
It may have limitations for projects that
- Use alternative languages
- Use Lua
- Non programming related projects

## Training and evaluation data

N/A

## Training procedure

Trained on 2x NVIDIA RTX 4090s

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 0.0003
- train_batch_size: 2
- eval_batch_size: 2
- seed: 42
- distributed_type: multi-GPU
- num_devices: 2
- gradient_accumulation_steps: 2
- total_train_batch_size: 8
- 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: 10
- training_steps: 105

### Training results

| Training Loss | Epoch  | Step | Validation Loss | Ppl    | Active (gib) | Allocated (gib) | Reserved (gib) |
|:-------------:|:------:|:----:|:---------------:|:------:|:------------:|:---------------:|:--------------:|
| No log        | 0      | 0    | 3.9969          | 54.428 | 11.21        | 11.1            | 12.26          |
| No log        | 0.2535 | 9    | nan             | nan    | 14.12        | 14.01           | 15.56          |
| 12.4054       | 0.5070 | 18   | nan             | nan    | 14.12        | 14.01           | 14.69          |
| 0.0           | 0.7606 | 27   | nan             | nan    | 14.12        | 14.01           | 14.69          |
| 0.0           | 1.0    | 36   | nan             | nan    | 14.12        | 14.01           | 14.69          |
| 0.0           | 1.2535 | 45   | nan             | nan    | 14.12        | 14.01           | 14.69          |
| 0.0           | 1.5070 | 54   | nan             | nan    | 14.12        | 14.01           | 14.69          |
| 0.0           | 1.7606 | 63   | nan             | nan    | 14.12        | 14.01           | 14.69          |
| 0.0           | 2.0    | 72   | nan             | nan    | 14.12        | 14.01           | 14.69          |
| 0.0           | 2.2535 | 81   | nan             | nan    | 14.12        | 14.01           | 14.69          |
| 0.0           | 2.5070 | 90   | nan             | nan    | 11.83        | 11.72           | 14.69          |
| 0.0           | 2.7606 | 99   | nan             | nan    | 14.12        | 14.01           | 14.69          |


### Framework versions

- PEFT 0.18.0
- Transformers 4.57.1
- Pytorch 2.8.0+cu128
- Datasets 4.4.1
- Tokenizers 0.22.1