axolotl version: 0.3.0

base_model: deepseek-ai/deepseek-coder-1.3b-base
model_type: AutoModelForCausalLM
trust_remote_code: true
load_in_8bit: false
load_in_4bit: false
strict: false


datasets:
  - path: CodeGPTPlus/typescript-0-500000-seq1024
    type: completion
    field: text


val_set_size: 0.001
output_dir:  ./fft-out

sequence_len: 1024

adapter:
lora_model_dir:
lora_r: 
lora_alpha: 
lora_dropout: 
lora_target_linear: 
lora_fan_in_fan_out:
lora_modules_to_save:

wandb_project: deepseek_1.3_fft
wandb_entity:
wandb_watch:
wandb_name: aws_a10g
wandb_log_model: end


gradient_accumulation_steps: 2
micro_batch_size: 20
num_epochs: 1
optimizer: adamw_bnb_8bit
adam_beta1: 0.9
adam_beta2: 0.999
adam_epsilon: 0.000001
max_grad_norm: 1.0
weight_decay: 0.1
lr_scheduler: cosine
learning_rate: 0.00002
train_on_inputs: false
group_by_length: false
bf16: true
fp16: false
tf32: false
gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

loss_watchdog_threshold: 5.0
loss_watchdog_patience: 3

hub_model_id: CodeGPTPlus/deepseek_coder_1.3b_typescript
hub_strategy: every_save
warmup_ratio: 0.01
evals_per_epoch: 20
saves_per_epoch: 3
debug:
deepspeed:

fsdp:
fsdp_config:
special_tokens:
  bos_token: "<｜begin▁of▁sentence｜>"
  eos_token: "<｜end▁of▁sentence｜>"
  pad_token: "<｜end▁of▁sentence｜>"

deepseek-coder-1.3b-typescript

CodeGPTPlus/deepseek-coder-1.3b-typescript, emerges as a fine-tuned iteration of deepseek-ai/deepseek-coder-1.3b-base, meticulously crafted by the CodeGPT team to excel in generating expert code in TypeScript. With specific fine-tuning for TypeScript and a dataset of 0.5B tokens, this model excels in producing precise and efficient solutions in this programming language.

The 16K window size and an additional fill-in-the-middle task are employed to deliver project-level code completion.

This new model stands as the ideal choice for those seeking a specialized code generator for TypeScript, backed by the expertise of the CodeGPT team.

It achieves the following results on the evaluation set:

• Loss: 0.7681

Model Developers CodeGPT Team

Variations 1.3B

Input Models input text only.

Output Models generate text only.

How to Use

This model is for completion purposes only. Here give some examples of how to use the model.

Running the model on a GPU

from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("CodeGPTPlus/deepseek-coder-1.3b-typescript",
                                          trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("CodeGPTPlus/deepseek-coder-1.3b-typescript",
                                              trust_remote_code=True).cuda()

input_text = """<｜fim▁begin｜>function quickSort(arr: number[]): number[] {
  if (arr.length <= 1) {
    return arr;
  }
  const pivot = arr[0];
  const left = [];
  const right = [];
<｜fim▁hole｜>
  return [...quickSort(left), pivot, ...quickSort(right)];
}<｜fim▁end｜>"""

inputs = tokenizer(input_text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_length=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Running with Ollama

Model: https://ollama.ai/codegpt/deepseek-coder-1.3b-typescript

ollama run codegpt/deepseek-coder-1.3b-typescript

Running with Ollama and CodeGPT Autocomplete in VSCode

Documentation: https://docs.codegpt.co/docs/tutorial-features/code_autocompletion

Select "Ollama - codegpt/deepseek-coder-1.3b-typescript" in the autocomplete model selector.

Then, write any code or comment in the vscode text editor, and the model will provide you with code suggestions through the CodeGPT code autocomplete.

Fill In the Middle (FIM)

<｜fim▁begin｜>function quickSort(arr: number[]): number[] {
  if (arr.length <= 1) {
    return arr;
  }
  const pivot = arr[0];
  const left = [];
  const right = [];
<｜fim▁hole｜>
  return [...quickSort(left), pivot, ...quickSort(right)];
}<｜fim▁end｜>

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 20
• eval_batch_size: 20
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 40
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-06
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 261
• num_epochs: 1

Training results

Framework versions

• Transformers 4.37.0.dev0
• Pytorch 2.0.1+cu118
• Datasets 2.16.1
• Tokenizers 0.15.0