metadata
language:
- code
license: bigcode-openrail-m
datasets:
- bigcode/the-stack-dedup
- Vipitis/Shadertoys-fine
pipeline_tag: text-generation
tags:
- code
- shader
base_model: bigcode/santacoder
widget:
- text: void mainImage( out vec4 fragColor, in vec2 fragCoord )
example_title: mainImage
group: Shadertoy
model-index:
- name: santacoder-finetuned-the-stack-glsl
results:
- task:
type: text-generation
name: ShaderEval
dataset:
type: Vipitis/Shadertoys-fine
name: Shadertoys-fine
config: return_completion
revision: 0.0.2
metrics:
- type: exact_match
value: 0.567
name: 300 samples, greedy decoding
verified: false
- type: exact_match
value: 0.59749
name: all samples, greedy decoding
verified: false
Santacoder finetuned on Shadertoys-fine for 1000 steps with a batch size of 2 and full sequence length of 2048. adapted finetuning script found here
Try model in the ShaderCoder demo space
Finetuning parameters
python3 train.py --model_path "bigcode/santacoder" \
--dataset_name "Vipitis/Shadertoys-fine" \
--data_column "code" \
--split "train" \
--seq_length 2048 \
--max_steps 1000 \
--batch_size 2 \
--gradient_accumulation_steps 4 \
--learning_rate 5e-5 \
--num_warmup_steps 100 \
--eval_freq 100 \
--save_freq 100 \
--log_freq 1 \
--output_dir "checkpoint_dir" \
--no_fp16
Main purpose of this model is to explore if finetuning models improves performance on ShaderEval, which reached 0.567 with 300 samples and 0.59749 on all samples.
Disclaimer
While the train/test split is held out, there is a lot of data contamination. The model results can't be trusted for this simple benchmark. Better tasks for the benchmark will be developed and tested against these models.
License carried over from model, however training data has an undefied license. Check details in Shadertoys.