Commit
•
11b49f8
1
Parent(s):
12e5537
Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,120 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: llama3
|
| 3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: llama3
|
| 3 |
+
datasets:
|
| 4 |
+
- BAAI/Infinity-Instruct
|
| 5 |
+
base_model:
|
| 6 |
+
- meta-llama/Meta-Llama-3.1-8B-Instruct
|
| 7 |
+
---
|
| 8 |
+
|
| 9 |
+
We prune the Llama-3.1-8B-Instruct to 1.4B and fine-tune it with LLM-Neo method，which combines LoRA and KD in one. Training data is sampling from BAAI/Infinity-Instruct for 1 Million lines.
|
| 10 |
+
|
| 11 |
+
## Benchmarks
|
| 12 |
+
|
| 13 |
+
In this section, we report the results for Llama3.1-Neo-1B-100w on standard automatic benchmarks. For all the evaluations, we use our internal evaluations library.
|
| 14 |
+
|
| 15 |
+
### Evaluation results
|
| 16 |
+
|
| 17 |
+
<table>
|
| 18 |
+
<tr>
|
| 19 |
+
<td><strong>Category</strong>
|
| 20 |
+
</td>
|
| 21 |
+
<td><strong>Benchmark</strong>
|
| 22 |
+
</td>
|
| 23 |
+
<td><strong>Version</strong>
|
| 24 |
+
</td>
|
| 25 |
+
<td><strong>n-shot</strong>
|
| 26 |
+
</td>
|
| 27 |
+
<td><strong>Metric</strong>
|
| 28 |
+
</td>
|
| 29 |
+
<td><strong>Value</strong>
|
| 30 |
+
</td>
|
| 31 |
+
<td><strong>Stderr</strong>
|
| 32 |
+
</td>
|
| 33 |
+
</tr>
|
| 34 |
+
<tr>
|
| 35 |
+
<td rowspan="2" >ARC
|
| 36 |
+
</td>
|
| 37 |
+
<td>ARC-Challenge</td>
|
| 38 |
+
<td>1</td>
|
| 39 |
+
<td>0</td>
|
| 40 |
+
<td>acc</td>
|
| 41 |
+
<td>0.1920</td>
|
| 42 |
+
<td>± 0.0115</td>
|
| 43 |
+
</tr>
|
| 44 |
+
<tr>
|
| 45 |
+
<td>ARC-Easy</td>
|
| 46 |
+
<td>1</td>
|
| 47 |
+
<td>0</td>
|
| 48 |
+
<td>acc</td>
|
| 49 |
+
<td>0.3834</td>
|
| 50 |
+
<td>± 0.0100</td>
|
| 51 |
+
</tr>
|
| 52 |
+
<tr>
|
| 53 |
+
<td rowspan="3" >CEVAL</td>
|
| 54 |
+
<td>CEVAL (valid)</td>
|
| 55 |
+
<td>N/A</td>
|
| 56 |
+
<td>0</td>
|
| 57 |
+
<td>acc</td>
|
| 58 |
+
<td>0.2370</td>
|
| 59 |
+
<td>± 0.0117</td>
|
| 60 |
+
</tr>
|
| 61 |
+
<tr>
|
| 62 |
+
<td>CEVAL (Accountant)</td>
|
| 63 |
+
<td>1</td>
|
| 64 |
+
<td>0</td>
|
| 65 |
+
<td>acc</td>
|
| 66 |
+
<td>0.2449</td>
|
| 67 |
+
<td>± 0.0621</td>
|
| 68 |
+
</tr>
|
| 69 |
+
<tr>
|
| 70 |
+
<td>CEVAL (Advanced Mathematics)</td>
|
| 71 |
+
<td>1</td>
|
| 72 |
+
<td>0</td>
|
| 73 |
+
<td>acc</td>
|
| 74 |
+
<td>0.3158</td>
|
| 75 |
+
<td>± 0.1096</td>
|
| 76 |
+
</tr>
|
| 77 |
+
<tr>
|
| 78 |
+
<td rowspan="2" >MMLU</td>
|
| 79 |
+
<td>MMLU</td>
|
| 80 |
+
<td>N/A</td>
|
| 81 |
+
<td>0</td>
|
| 82 |
+
<td>acc</td>
|
| 83 |
+
<td>0.2439</td>
|
| 84 |
+
<td>± 0.0036</td>
|
| 85 |
+
</tr>
|
| 86 |
+
<tr>
|
| 87 |
+
<td>MMLU (Abstract Algebra)</td>
|
| 88 |
+
<td>0</td>
|
| 89 |
+
<td>0</td>
|
| 90 |
+
<td>acc</td>
|
| 91 |
+
<td>0.2500</td>
|
| 92 |
+
<td>± 0.0435</td>
|
| 93 |
+
</tr>
|
| 94 |
+
<tr>
|
| 95 |
+
<td rowspan="2" >PIQA</td>
|
| 96 |
+
<td>PIQA</td>
|
| 97 |
+
<td>1</td>
|
| 98 |
+
<td>0</td>
|
| 99 |
+
<td>acc</td>
|
| 100 |
+
<td>0.5843</td>
|
| 101 |
+
<td>± 0.0115</td>
|
| 102 |
+
</tr>
|
| 103 |
+
<tr>
|
| 104 |
+
<td>PIQA (Normalized)</td>
|
| 105 |
+
<td>1</td>
|
| 106 |
+
<td>0</td>
|
| 107 |
+
<td>acc_norm</td>
|
| 108 |
+
<td>0.5822</td>
|
| 109 |
+
<td>± 0.0115</td>
|
| 110 |
+
</tr>
|
| 111 |
+
<tr>
|
| 112 |
+
<td>Winogrande</td>
|
| 113 |
+
<td>Winogrande</td>
|
| 114 |
+
<td>1</td>
|
| 115 |
+
<td>0</td>
|
| 116 |
+
<td>acc</td>
|
| 117 |
+
<td>0.5249</td>
|
| 118 |
+
<td>± 0.0140</td>
|
| 119 |
+
</tr>
|
| 120 |
+
</table>
|