cosmo3769's picture
Create README.md
1ac95e3 verified
# starcoderbase-1b-GPTQ
Quantized starcoderbase-1b model to GPTQ format (4-bit precision) using [Auto-GPTQ](https://github.com/AutoGPTQ/AutoGPTQ).
[Quantization script](https://github.com/cosmo3769/Quantized-LLMs/blob/main/notebooks/quantize-starcoderbase-1b-4bit-gptq.ipynb)
## Benchmark
[Benchmarking script](https://github.com/cosmo3769/Quantized-LLMs/blob/main/notebooks/llmbenchmark-starcoderbase-1b-lm-eval-harness.ipynb)
### Baseline starcoderbase-1b model (non-quantized)
| Tasks |Version|Filter|n-shot| Metric |Value | |Stderr|
|-----------------------|-------|------|------|---------------|-----:|---|-----:|
|codexglue_code2text |N/A |none |None |smoothed_bleu_4|0.8767|± |0.0592|
| - code2text_go | 1|none |None |smoothed_bleu_4|1.0054|± |0.0983|
| - code2text_java | 1|none |None |smoothed_bleu_4|1.2158|± |0.1657|
| - code2text_javascript| 1|none |None |smoothed_bleu_4|0.8560|± |0.0429|
| - code2text_php | 1|none |None |smoothed_bleu_4|0.9879|± |0.0887|
| - code2text_python | 1|none |None |smoothed_bleu_4|1.1950|± |0.2819|
| - code2text_ruby | 3|none |None |smoothed_bleu_4|0.0000|± |0.0000|
| Groups |Version|Filter|n-shot| Metric |Value | |Stderr|
|-------------------|-------|------|------|---------------|-----:|---|-----:|
|codexglue_code2text|N/A |none |None |smoothed_bleu_4|0.8767|± |0.0592|
| Tasks |Version|Filter|n-shot| Metric |Value| |Stderr|
|---------------------------------------------|------:|------|------|-----------|----:|---|-----:|
|bigbench_code_line_description_generate_until| 1|none |None |exact_match| 0|± | 0|
| Tasks |Version|Filter|n-shot|Metric|Value| |Stderr|
|----------------------------------------------|------:|------|------|------|----:|---|-----:|
|bigbench_code_line_description_multiple_choice| 0|none |None |acc | 0.15|± |0.0465|
### Quantized starcoderbase-1b model to GPTQ format
| Tasks |Version|Filter|n-shot| Metric |Value | |Stderr|
|-----------------------|-------|------|------|---------------|-----:|---|-----:|
|codexglue_code2text |N/A |none |None |smoothed_bleu_4|0.7959|± |0.2180|
| - code2text_go | 1|none |None |smoothed_bleu_4|0.9280|± |0.0291|
| - code2text_java | 1|none |None |smoothed_bleu_4|1.2112|± |0.1703|
| - code2text_javascript| 1|none |None |smoothed_bleu_4|0.8848|± |0.0391|
| - code2text_php | 1|none |None |smoothed_bleu_4|0.6055|± |0.6055|
| - code2text_python | 1|none |None |smoothed_bleu_4|1.1460|± |1.1460|
| - code2text_ruby | 3|none |None |smoothed_bleu_4|0.0000|± |0.0000|
| Groups |Version|Filter|n-shot| Metric |Value | |Stderr|
|-------------------|-------|------|------|---------------|-----:|---|-----:|
|codexglue_code2text|N/A |none |None |smoothed_bleu_4|0.7959|± | 0.218|
| Tasks |Version|Filter|n-shot| Metric |Value| |Stderr|
|---------------------------------------------|------:|------|------|-----------|----:|---|-----:|
|bigbench_code_line_description_generate_until| 1|none |None |exact_match| 0|± | 0|
| Tasks |Version|Filter|n-shot|Metric|Value | |Stderr|
|----------------------------------------------|------:|------|------|------|-----:|---|-----:|
|bigbench_code_line_description_multiple_choice| 0|none |None |acc |0.1333|± |0.0443|