thegovind's picture
Split the longest sentences and fit the install command on screen (from 338f312) (part 2)
09fbc15 verified
Raw History Blame Contribute Delete
9.89 kB
<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<meta name="color-scheme" content="only light">
<meta name="theme-color" content="#f5f5f5">
<title>Results · Weight-space programming lab</title>
<meta name="description" content="&quot;Cut answers ending in repetition&quot; counts two kinds of ending among the answers cut at the cap. The first is a word-for-word loop (one block of words repeated back to back at least 3 times, over at…">
<link rel="icon" href="../static/favicon.svg" type="image/svg+xml">
<link rel="preload" href="../static/fonts/geist-latin-wght-normal.woff2" as="font" type="font/woff2" crossorigin>
<link rel="stylesheet" href="../static/site.css">
<script>document.documentElement.classList.add("js")</script>
<script defer src="../static/site.js"></script>
</head>
<body data-copy="Copy" data-copied="Copied">
<a class="skip" href="#main">Skip to content</a>
<header class="top">
<div class="top-in">
<a class="brand" href="../index.html">Weight-space programming lab</a>
<nav class="top-links" aria-label="Site"><a href="../index.html">Home</a><a href="../quickstart.html">Quick start</a><a href="../index.html#the-nine-labs">Labs</a><a href="../reference/theory-guide.html">Theory</a><a href="summary.html" aria-current="page">Results</a><a href="https://huggingface.co/spaces/thegovind/weight-space-programming/tree/main/code/" target="_blank" rel="noopener">Code</a></nav>
</div>
</header>
<div class="page solo">
<main id="main" class="doc">
<article class="prose">
<h1 id="results">Results</h1>
<div class="table-wrap"><table>
<thead>
<tr>
<th>run</th>
<th>problems × samples</th>
<th>accuracy %</th>
<th>cut at the cap %</th>
<th>mean tokens</th>
<th>correct when finished %</th>
<th>cut answers ending in repetition %</th>
<th>gain vs baseline (pp)</th>
<th>rescued / lost problems</th>
</tr>
</thead>
<tbody>
<tr>
<td>bench/base/gsm8k</td>
<td>1319×2</td>
<td>14.2 ± 0.8</td>
<td>87.7</td>
<td>29,370</td>
<td>97.5</td>
<td>98.3</td>
<td>-</td>
<td>-</td>
</tr>
<tr>
<td>bench/base/ifeval</td>
<td>541×2</td>
<td>21.2 ± 1.4</td>
<td>73.5</td>
<td>24,976</td>
<td>78.7</td>
<td>96.1</td>
<td>-</td>
<td>-</td>
</tr>
<tr>
<td>bench/base/math500</td>
<td>500×4</td>
<td>7.3 ± 0.9</td>
<td>93.3</td>
<td>30,956</td>
<td>100.0</td>
<td>96.2</td>
<td>-</td>
<td>-</td>
</tr>
<tr>
<td>bench/p50c/gsm8k</td>
<td>1319×2</td>
<td>66.9 ± 1.2</td>
<td>6.3</td>
<td>11,742</td>
<td>71.3</td>
<td>29.7</td>
<td>+52.7 ± 1.1</td>
<td>711 / 2</td>
</tr>
<tr>
<td>bench/p50c/ifeval</td>
<td>541×2</td>
<td>63.4 ± 1.8</td>
<td>1.4</td>
<td>3,984</td>
<td>64.3</td>
<td>80.0</td>
<td>+42.2 ± 2.0</td>
<td>261 / 22</td>
</tr>
<tr>
<td>bench/p50c/math500</td>
<td>500×4</td>
<td>51.8 ± 1.9</td>
<td>31.0</td>
<td>21,950</td>
<td>75.0</td>
<td>34.7</td>
<td>+44.5 ± 1.7</td>
<td>263 / 0</td>
</tr>
<tr>
<td>build/base</td>
<td>128×2</td>
<td>11.7 ± 2.5</td>
<td>89.8</td>
<td>29,872</td>
<td>100.0</td>
<td>96.5</td>
<td>-</td>
<td>-</td>
</tr>
<tr>
<td>dev/a-4</td>
<td>256×2</td>
<td>19.7 ± 1.9</td>
<td>65.8</td>
<td>26,063</td>
<td>55.4</td>
<td>96.4</td>
<td>+7.6 ± 2.2</td>
<td>54 / 24</td>
</tr>
<tr>
<td>dev/a1</td>
<td>256×2</td>
<td>22.7 ± 2.1</td>
<td>77.1</td>
<td>26,964</td>
<td>96.6</td>
<td>93.9</td>
<td>+10.5 ± 1.8</td>
<td>42 / 7</td>
</tr>
<tr>
<td>dev/a12</td>
<td>256×2</td>
<td>51.8 ± 2.8</td>
<td>12.3</td>
<td>13,515</td>
<td>59.0</td>
<td>9.5</td>
<td>+39.6 ± 2.7</td>
<td>102 / 0</td>
</tr>
<tr>
<td>dev/a2</td>
<td>256×2</td>
<td>35.2 ± 2.6</td>
<td>62.7</td>
<td>23,911</td>
<td>92.1</td>
<td>94.7</td>
<td>+23.0 ± 2.3</td>
<td>71 / 5</td>
</tr>
<tr>
<td>dev/a4</td>
<td>256×2</td>
<td>59.2 ± 2.8</td>
<td>29.1</td>
<td>17,730</td>
<td>82.9</td>
<td>80.5</td>
<td>+47.1 ± 2.6</td>
<td>118 / 0</td>
</tr>
<tr>
<td>dev/a8</td>
<td>256×2</td>
<td>60.7 ± 2.8</td>
<td>14.1</td>
<td>14,589</td>
<td>70.7</td>
<td>26.4</td>
<td>+48.6 ± 2.6</td>
<td>119 / 0</td>
</tr>
<tr>
<td>dev/all1</td>
<td>256×2</td>
<td>28.7 ± 2.3</td>
<td>71.9</td>
<td>25,609</td>
<td>97.9</td>
<td>95.7</td>
<td>+16.6 ± 2.0</td>
<td>61 / 6</td>
</tr>
<tr>
<td>dev/all2</td>
<td>256×2</td>
<td>49.4 ± 2.8</td>
<td>46.7</td>
<td>20,117</td>
<td>91.9</td>
<td>87.0</td>
<td>+37.3 ± 2.5</td>
<td>97 / 0</td>
</tr>
<tr>
<td>dev/base</td>
<td>256×2</td>
<td>12.1 ± 1.6</td>
<td>88.7</td>
<td>29,683</td>
<td>100.0</td>
<td>96.9</td>
<td>-</td>
<td>-</td>
</tr>
<tr>
<td>dev/base+max4k</td>
<td>256×2</td>
<td>5.9 ± 1.1</td>
<td>95.5</td>
<td>4,019</td>
<td>100.0</td>
<td>18.2</td>
<td>-6.2 ± 1.2</td>
<td>3 / 30</td>
</tr>
<tr>
<td>dev/base+nothink</td>
<td>256×2</td>
<td>52.1 ± 2.7</td>
<td>18.0</td>
<td>2,270</td>
<td>63.6</td>
<td>7.6</td>
<td>+40.0 ± 2.5</td>
<td>115 / 2</td>
</tr>
<tr>
<td>dev/base+pp0</td>
<td>256×2</td>
<td>4.5 ± 0.9</td>
<td>99.0</td>
<td>32,472</td>
<td>100.0</td>
<td>98.4</td>
<td>-7.6 ± 1.6</td>
<td>11 / 42</td>
</tr>
<tr>
<td>dev/base+pp2</td>
<td>256×2</td>
<td>16.2 ± 1.9</td>
<td>84.6</td>
<td>28,614</td>
<td>98.7</td>
<td>96.8</td>
<td>+4.1 ± 1.7</td>
<td>24 / 13</td>
</tr>
<tr>
<td>dev/base+rep1.1</td>
<td>256×2</td>
<td>34.4 ± 2.6</td>
<td>61.7</td>
<td>23,301</td>
<td>89.8</td>
<td>88.3</td>
<td>+22.3 ± 2.2</td>
<td>62 / 3</td>
</tr>
<tr>
<td>dev/base+temp1.0</td>
<td>256×2</td>
<td>27.9 ± 2.5</td>
<td>71.3</td>
<td>27,430</td>
<td>95.9</td>
<td>59.5</td>
<td>+15.8 ± 2.1</td>
<td>52 / 9</td>
</tr>
<tr>
<td>dev/capA-a6</td>
<td>256×2</td>
<td>43.0 ± 2.6</td>
<td>52.5</td>
<td>22,699</td>
<td>89.3</td>
<td>81.4</td>
<td>+30.9 ± 2.3</td>
<td>95 / 1</td>
</tr>
<tr>
<td>dev/capA-alpha-8</td>
<td>256×2</td>
<td>42.8 ± 2.8</td>
<td>53.5</td>
<td>22,772</td>
<td>89.9</td>
<td>78.5</td>
<td>+30.7 ± 2.5</td>
<td>87 / 4</td>
</tr>
<tr>
<td>dev/capA-max-samples-16</td>
<td>256×2</td>
<td>17.6 ± 1.9</td>
<td>82.6</td>
<td>28,525</td>
<td>96.6</td>
<td>96.0</td>
<td>+5.5 ± 1.9</td>
<td>34 / 17</td>
</tr>
<tr>
<td>dev/capB-a6</td>
<td>256×2</td>
<td>58.2 ± 2.9</td>
<td>30.7</td>
<td>19,899</td>
<td>83.7</td>
<td>49.0</td>
<td>+46.1 ± 2.6</td>
<td>113 / 0</td>
</tr>
<tr>
<td>dev/capB-alpha-8</td>
<td>256×2</td>
<td>54.5 ± 2.9</td>
<td>34.0</td>
<td>20,870</td>
<td>81.7</td>
<td>37.9</td>
<td>+42.4 ± 2.6</td>
<td>107 / 1</td>
</tr>
<tr>
<td>dev/capB-max-samples-16</td>
<td>256×2</td>
<td>53.7 ± 2.8</td>
<td>39.6</td>
<td>20,725</td>
<td>88.7</td>
<td>76.8</td>
<td>+41.6 ± 2.6</td>
<td>110 / 0</td>
</tr>
<tr>
<td>dev/l15only</td>
<td>256×2</td>
<td>60.4 ± 2.9</td>
<td>28.5</td>
<td>17,362</td>
<td>84.2</td>
<td>67.8</td>
<td>+48.2 ± 2.7</td>
<td>117 / 0</td>
</tr>
<tr>
<td>dev/l19only</td>
<td>256×2</td>
<td>46.9 ± 2.7</td>
<td>49.0</td>
<td>20,729</td>
<td>91.2</td>
<td>90.0</td>
<td>+34.8 ± 2.4</td>
<td>100 / 0</td>
</tr>
<tr>
<td>dev/p50c</td>
<td>256×2</td>
<td>61.5 ± 2.8</td>
<td>19.7</td>
<td>16,104</td>
<td>76.6</td>
<td>24.8</td>
<td>+49.4 ± 2.6</td>
<td>125 / 0</td>
</tr>
<tr>
<td>dev/p50c+nothink</td>
<td>256×2</td>
<td>53.5 ± 2.8</td>
<td>16.4</td>
<td>2,306</td>
<td>64.0</td>
<td>0.0</td>
<td>+1.4 ± 1.9</td>
<td>22 / 26</td>
</tr>
<tr>
<td>dev/randcols</td>
<td>256×2</td>
<td>14.6 ± 1.9</td>
<td>86.7</td>
<td>29,113</td>
<td>100.0</td>
<td>97.5</td>
<td>+2.5 ± 1.6</td>
<td>23 / 20</td>
</tr>
<tr>
<td>dev/randdir</td>
<td>256×2</td>
<td>14.5 ± 1.8</td>
<td>87.3</td>
<td>29,486</td>
<td>98.5</td>
<td>97.3</td>
<td>+2.3 ± 1.6</td>
<td>22 / 18</td>
</tr>
<tr>
<td>dev/rebuilt</td>
<td>256×2</td>
<td>61.5 ± 2.8</td>
<td>17.6</td>
<td>16,350</td>
<td>74.6</td>
<td>32.2</td>
<td>+49.4 ± 2.6</td>
<td>122 / 0</td>
</tr>
<tr>
<td>dev/rebuilt-layers15-19-21</td>
<td>256×2</td>
<td>61.3 ± 2.8</td>
<td>21.1</td>
<td>15,947</td>
<td>77.5</td>
<td>38.0</td>
<td>+49.2 ± 2.6</td>
<td>122 / 0</td>
</tr>
<tr>
<td>dev/rebuilt-pairs2</td>
<td>256×2</td>
<td>62.3 ± 2.7</td>
<td>25.2</td>
<td>16,080</td>
<td>83.0</td>
<td>74.4</td>
<td>+50.2 ± 2.6</td>
<td>129 / 0</td>
</tr>
<tr>
<td>dev/rebuilt-windows36-36</td>
<td>256×2</td>
<td>61.9 ± 2.8</td>
<td>21.3</td>
<td>16,141</td>
<td>78.7</td>
<td>35.8</td>
<td>+49.8 ± 2.7</td>
<td>122 / 0</td>
</tr>
<tr>
<td>quick/base</td>
<td>16×2</td>
<td>18.8 ± 7.7</td>
<td>84.4</td>
<td>28,375</td>
<td>100.0</td>
<td>100.0</td>
<td>-</td>
<td>-</td>
</tr>
<tr>
<td>quick/p50c</td>
<td>16×2</td>
<td>65.6 ± 11.8</td>
<td>9.4</td>
<td>11,664</td>
<td>72.4</td>
<td>0.0</td>
<td>+46.9 ± 11.6</td>
<td>6 / 0</td>
</tr>
</tbody>
</table></div>
<p>&quot;Cut answers ending in repetition&quot; counts two kinds of ending among the answers cut at the cap. The first is a word-for-word loop (one block of words repeated back to back at least 3 times, over at least 40 words). The second is a looser circling (at least half of the 4-word phrases in the last 200 words occur 3 or more times there). See verbatim_loop_end and repetitive_end in wsp/grading.py.</p>
</article>
</main>
</div>
<footer class="foot">
<div class="foot-in">
<p class="foot-note">Based on Experiment 8 of "<a href="https://rikiyat.github.io/blog/weight-space-programming/" target="_blank" rel="noopener">Sparks of Weight-Space Programming</a>". Logs and the agent's texts come from <a href="https://github.com/RikiyaT/lm_understand-logs" target="_blank" rel="noopener">RikiyaT/lm_understand-logs</a>, <a href="https://creativecommons.org/licenses/by/4.0/" target="_blank" rel="noopener">CC BY 4.0</a>.</p>
<p class="foot-links"><a href="https://huggingface.co/spaces/thegovind/weight-space-programming/blob/main/code/results/summary.md" target="_blank" rel="noopener">View source</a></p>
</div>
</footer>
</body>
</html>