base_model: deepseek-ai/deepseek-coder-6.7b-instruct
hub_model_id: darwinkernelpanic/deepseek-coder-6.7b-instruct-luau
hub_strategy: end
trust_remote_code: true

load_in_8bit: false
load_in_4bit: true

datasets:
  - path: darwinkernelpanic/luau_corpus_axolotl
    type: completion
    field_instruction: prompt
    field_output: completion

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

sequence_len: 3072
sample_packing: true
eval_sample_packing: true

adapter: qlora
lora_model_dir:
lora_r: 32
lora_alpha: 32
lora_dropout: 0.05
lora_target_linear: true

wandb_project: deepseek-luau-finetune
wandb_entity:
wandb_watch:
wandb_name: deepseek-coder-6.7b-luau
wandb_log_model:

gradient_accumulation_steps: 2
micro_batch_size: 6
num_epochs: 3
optimizer: adamw_torch_fused
lr_scheduler: cosine
learning_rate: 0.0002
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: []
fsdp_config: {}

special_tokens:
  pad_token: "<|EOT|>"

deepseek-coder-6.7b-instruct-luau

This model is a fine-tuned version of deepseek-ai/deepseek-coder-6.7b-instruct on the darwinkernelpanic/luau_corpus_axolotl dataset. It achieves the following results on the evaluation set:

Loss: 1.6346
Ppl: 5.1272
Memory/max Active (gib): 10.65
Memory/max Allocated (gib): 10.65
Memory/device Reserved (gib): 11.93

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), 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 1x RTX 6000Ada

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0002
train_batch_size: 6
eval_batch_size: 6
seed: 42
gradient_accumulation_steps: 2
total_train_batch_size: 12
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: 16
training_steps: 162

Training results

Training Loss	Epoch	Step	Validation Loss	Ppl	Active (gib)	Allocated (gib)	Reserved (gib)
No log	0	0	3.8515	47.0637	7.0	7.0	7.26
3.2644	0.2593	14	2.8645	17.5407	10.65	10.65	12.22
2.6242	0.5185	28	2.2633	9.6147	12.27	12.27	14.58
2.0431	0.7778	42	2.0479	7.7515	10.65	10.65	13.92
1.9054	1.0370	56	1.9163	6.796	10.65	10.65	14.72
1.7318	1.2963	70	1.8184	6.1622	7.61	7.61	13.92
1.6119	1.5556	84	1.7550	5.7836	12.27	12.27	14.54
1.6022	1.8148	98	1.7048	5.5006	10.65	10.65	14.23
1.6249	2.0741	112	1.6723	5.3242	10.65	10.65	13.99
1.4995	2.3333	126	1.6503	5.2088	10.65	10.65	11.93
1.4803	2.5926	140	1.6381	5.1452	7.61	7.61	14.58
1.4872	2.8519	154	1.6346	5.1272	10.65	10.65	11.93

Framework versions

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