XL-LuaCopilot-1.7B-FFT is a large language model (LLM) based on the Qwen architecture(Qwen3-1.7B), specifically designed for code generation tasks in Lua programming language. It has been full fine-tuned (FFT) to improve its performance and efficiency when generating Lua code.

I sugggest you use "chat_template_kwargs": {"enable_thinking": false} because my train data with none thinking. I also found low temperature ususually works well for code generation tasks.

checkpoint-37000 is the checkpoint where the model had just entered the plateau phase with a lower loss, and it might be better than checkpoint-46000. However, the GGUF files I provide will still be released based on the final checkpoint at the end of training.

Train Samples

1184000 (steps=37000)x(per_device_train_batch_size=8)x(gradient_accumulation_steps=4)x(device=1)

How To Use

With OpenAI Compatible API (llama.cpp:llama-server)

json

-> REQUEST ->

{
    "model": "XL-LuaCopilot-1.7B-FFT",
    "messages": [
        {"role": "system","content": "prefix"},
        {"role": "user","content": "do\n--打印：你好世界\n  local tex"},
        {"role": "system","content": "suffix"},
        {"role": "user","content": "nd"},
        {"role": "system","content": "middle"}
    ],
    "stream": false,
    "cache_prompt": false,
    "samplers": "edkypmxt",
    "temperature": 0.2,
    "dynatemp_range": 0.1,
    "dynatemp_exponent": 1,
    "top_k": 20,
    "top_p": 0.9,
    "min_p": 0.05,
    "typical_p": 1,
    "xtc_probability": 0,
    "xtc_threshold": 0.1,
    "repeat_last_n": 32,
    "repeat_penalty": 1.1,
    "presence_penalty": 0,
    "frequency_penalty": 0.5,
    "dry_multiplier": 0,
    "dry_base": 1.75,
    "dry_allowed_length": 2,
    "dry_penalty_last_n": -1,
    "max_tokens": -1,
    "timings_per_token": true,
    "chat_template_kwargs": {"enable_thinking": false}
}

-> RESPONSE ->

{
    "choices": [
        {
            "finish_reason": "stop",
            "index": 0,
            "message": {
                "role": "assistant",
                "content": "<think>\n\n</think>\n\nt = \"你好世界\"\n  print(text)\ne"
            }
        }
    ],
    ...
}

I know Qwen has <|fim_prefix|> / <|fim_suffix|> / <|fim_middle|> tokens, but I'm not sure Qwen3 trains these tokens (I just know Qwen2.5-Coder does). To use code generation easily, I use chatml format.

If you just want to chat with it, you can use some tricks like this:

<|im_end|>
<|im_start|>system
prefix<|im_end|>
<|im_start|>user
do
    --打印：你好世界
    local tex<|im_end|>
<|im_start|>system
suffix<|im_end|>
<|im_start|>user
nd<|im_end|>
<|im_start|>system
middle

It dosen't work very well, but it's a good way let you fast try. It will convert to this prompt text:

<|im_start|>user
<|im_end|>
<|im_start|>system
prefix<|im_end|>
<|im_start|>user
do
    --打印：你好世界
    local tex<|im_end|>
<|im_start|>system
suffix<|im_end|>
<|im_start|>user
nd<|im_end|>
<|im_start|>system
middle<|im_end|>

Hope model skip first <|im_start|>user\n<|im_end|> part.

Train Device

Online GPU is Expensive !

类别	配置详情
镜像	Ubuntu 22.04
PyTorch	2.5.1
Python	3.12
CUDA	12.4
GPU	RTX 4090 (24GB) * 1
CPU	25 vCPU Intel(R) Xeon(R) Platinum 8481C
内存	90GB
硬盘	30 GB + 50 GB
时长	3 Day

nwdxlgzs/XL-LuaCopilot-1.7B-FFT-checkpoint-37000

XL-LuaCopilot-1.7B-FFT

Train Samples

How To Use

Train Device

nwdxlgzs/
XL-LuaCopilot-1.7B-FFT-checkpoint-37000