Added PPL results with new-eval
Browse files
README.md
CHANGED
|
@@ -40,12 +40,30 @@ Refs:
|
|
| 40 |
|
| 41 |
# Benchmarks
|
| 42 |
|
| 43 |
-
Best results in **bold**.
|
| 44 |
|
| 45 |
-
|
|
|
|
|
|
|
| 46 |
|---|---|---|---|---|
|
| 47 |
-
| 4bit-GPTQ - TheBloke/vicuna-13B-1.1-GPTQ-4bit-128g | 8.517391204833984 | **7.058407783508301** |
|
| 48 |
-
| 8bit-GPTQ - Thireus/Vicuna13B-v1.1-8bit-128g | **8.508771896362305** |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 49 |
|
| 50 |
<br>
|
| 51 |
<br>
|
|
|
|
| 40 |
|
| 41 |
# Benchmarks
|
| 42 |
|
| 43 |
+
Using https://github.com/qwopqwop200/GPTQ-for-LLaMa/. Best results in **bold**.
|
| 44 |
|
| 45 |
+
*`--benchmark 2048 --check` results:*
|
| 46 |
+
|
| 47 |
+
| Model | wikitext2 PPL | ptb PPL | c4 PPL | VRAM Utilization |
|
| 48 |
|---|---|---|---|---|
|
| 49 |
+
| 4bit-GPTQ - TheBloke/vicuna-13B-1.1-GPTQ-4bit-128g | 8.517391204833984 | 20.888103485107422 | **7.058407783508301** | **8670.26953125** |
|
| 50 |
+
| 8bit-GPTQ - Thireus/Vicuna13B-v1.1-8bit-128g | **8.508771896362305** | **20.75649070739746** | 7.105874538421631 | 14840.26171875 |
|
| 51 |
+
|
| 52 |
+
*`--eval` results (pending):*
|
| 53 |
+
|
| 54 |
+
| Model | wikitext2 PPL | ptb PPL | c4 PPL |
|
| 55 |
+
|---|---|---|---|
|
| 56 |
+
| 4bit-GPTQ - TheBloke/vicuna-13B-1.1-GPTQ-4bit-128g | | | |
|
| 57 |
+
| 8bit-GPTQ - Thireus/Vicuna13B-v1.1-8bit-128g | | | |
|
| 58 |
+
|
| 59 |
+
*`--new-eval --eval` results:*
|
| 60 |
+
|
| 61 |
+
| Model | wikitext2 PPL | ptb-new PPL | c4-new PPL |
|
| 62 |
+
|---|---|---|---|
|
| 63 |
+
| 4bit-GPTQ - TheBloke/vicuna-13B-1.1-GPTQ-4bit-128g | 7.119165420532227 | 35.637290954589844 | 9.550592422485352 |
|
| 64 |
+
| 8bit-GPTQ - Thireus/Vicuna13B-v1.1-8bit-128g | **6.988043308258057** | **34.264320373535156** | **9.426002502441406** |
|
| 65 |
+
|
| 66 |
+
PPL = Perplexity - https://huggingface.co/docs/transformers/perplexity
|
| 67 |
|
| 68 |
<br>
|
| 69 |
<br>
|