Codemind-study / m22 /eval_results.json
B4K2xx's picture
Add m22 weights and results
f296a89 verified
Raw
History Blame Contribute Delete
7.18 kB
{
"model_id": "m22",
"checkpoint": "/home/akshat/codemindv2/results/m22/final.pt",
"trained_step": 15258,
"total_params": 64017408,
"active_params": 64017408,
"flops_per_forward": 128034816.0,
"val_loss": 3.5585316568613052,
"val_perplexity": 35.11160339377922,
"bits_per_token": 5.133875974200191,
"ppl_tokens_evaluated": 655360,
"prefill_tok_per_sec": 195315.3982182166,
"inference_tok_per_sec": 326.0861900573855,
"inference_tok_per_sec_kvcache": null,
"kvcache_speedup": null,
"gen_tokens": 128,
"peak_infer_vram_gb": 0.394,
"peak_train_vram_gb": 4.093,
"closeness_answer_ppl": 33.765,
"closeness_hit_rate": 0.45,
"closeness_token_f1": 0.093,
"closeness_n": 20,
"closeness_items": [
{
"prompt": "The capital of France is",
"answer": " Paris",
"generated": " the capital of France.\n",
"answer_nll": 3.2356,
"answer_ppl": 25.421,
"hit": false,
"token_f1": 0.0
},
{
"prompt": "The capital of Japan is",
"answer": " Tokyo",
"generated": " the capital of the United States",
"answer_nll": 2.6915,
"answer_ppl": 14.754,
"hit": false,
"token_f1": 0.0
},
{
"prompt": "The largest planet in the solar system is",
"answer": " Jupiter",
"generated": " the Sun, which is the",
"answer_nll": 3.1711,
"answer_ppl": 23.834,
"hit": false,
"token_f1": 0.0
},
{
"prompt": "Water is made of hydrogen and",
"answer": " oxygen",
"generated": " oxygen.\n- The oxygen",
"answer_nll": 0.9604,
"answer_ppl": 2.613,
"hit": true,
"token_f1": 0.4
},
{
"prompt": "The opposite of hot is",
"answer": " cold",
"generated": " the coldest, the cold",
"answer_nll": 3.1722,
"answer_ppl": 23.86,
"hit": true,
"token_f1": 0.4
},
{
"prompt": "The opposite of up is",
"answer": " down",
"generated": " the fact that the two are",
"answer_nll": 6.9194,
"answer_ppl": 1011.664,
"hit": false,
"token_f1": 0.0
},
{
"prompt": "Two plus two equals",
"answer": " four",
"generated": ", and the two equals,",
"answer_nll": 5.168,
"answer_ppl": 175.56,
"hit": false,
"token_f1": 0.0
},
{
"prompt": "The sun rises in the",
"answer": " east",
"generated": " sky, and the sun rises",
"answer_nll": 3.1618,
"answer_ppl": 23.613,
"hit": false,
"token_f1": 0.0
},
{
"prompt": "The first president of the United States was George",
"answer": " Washington",
"generated": " Washington, who was the first",
"answer_nll": 0.9015,
"answer_ppl": 2.463,
"hit": true,
"token_f1": 0.0
},
{
"prompt": "The speed of light is approximately 300,000 kilometers per",
"answer": " second",
"generated": " second.\nThe speed of",
"answer_nll": 0.2886,
"answer_ppl": 1.335,
"hit": true,
"token_f1": 0.0
},
{
"prompt": "A triangle has three",
"answer": " sides",
"generated": " sides, and the sides are",
"answer_nll": 2.1492,
"answer_ppl": 8.578,
"hit": true,
"token_f1": 0.333
},
{
"prompt": "The chemical symbol for gold is",
"answer": " Au",
"generated": " the gold of the world.",
"answer_nll": 9.4926,
"answer_ppl": 13261.109,
"hit": false,
"token_f1": 0.0
},
{
"prompt": "The freezing point of water in Celsius is",
"answer": " zero",
"generated": " about 1.5 degrees Celsius",
"answer_nll": 5.8433,
"answer_ppl": 344.908,
"hit": false,
"token_f1": 0.0
},
{
"prompt": "The author of Romeo and Juliet is William",
"answer": " Shakespeare",
"generated": " Shakespeare.\nThe author of",
"answer_nll": 1.3143,
"answer_ppl": 3.722,
"hit": true,
"token_f1": 0.0
},
{
"prompt": "The Earth revolves around the",
"answer": " Sun",
"generated": " Sun, and the Sun revolves",
"answer_nll": 0.9692,
"answer_ppl": 2.636,
"hit": true,
"token_f1": 0.333
},
{
"prompt": "The plural of mouse is",
"answer": " mice",
"generated": " the plural of mouse.\n",
"answer_nll": 10.0879,
"answer_ppl": 24050.054,
"hit": false,
"token_f1": 0.0
},
{
"prompt": "Photosynthesis occurs in the",
"answer": " leaves",
"generated": " chlorophyll-rich waters",
"answer_nll": 3.7037,
"answer_ppl": 40.596,
"hit": false,
"token_f1": 0.0
},
{
"prompt": "The human heart pumps",
"answer": " blood",
"generated": " blood to the heart.\n",
"answer_nll": 1.0486,
"answer_ppl": 2.854,
"hit": true,
"token_f1": 0.4
},
{
"prompt": "The currency used in the United States is the",
"answer": " dollar",
"generated": " United States Dollar.\nThe",
"answer_nll": 3.1815,
"answer_ppl": 24.084,
"hit": true,
"token_f1": 0.0
},
{
"prompt": "The tallest animal in the world is the",
"answer": " giraffe",
"generated": " giant elephant.\nThe elephant",
"answer_nll": 2.928,
"answer_ppl": 18.69,
"hit": false,
"token_f1": 0.0
}
],
"generation_samples": {
"The most important thing about sleep is": "The most important thing about sleep is waking up. It can be easy to go back to sleep for a while. So, it is important to give proper care to your sleep.\nIn some cases sleep is disrupted because of the lack of a proper sleep environment. If the body doesn\u2019t get enough sleep, your body can",
"Scientists have discovered that": "Scientists have discovered that when their muscles contract, it can trigger a reaction. This is how the muscles contract when they're stressed.\nTrying to work on the muscles in these situations is the best way to help your heart and brain function.\nYour body is only a little bit smaller than any other organs.\n",
"In the history of mathematics,": "In the history of mathematics, we had very much in this field of science a year before, we have only two problems which have always proved to us quite difficult.\nI shall attempt to describe our theory of an elementary school which we have only to discuss in terms of mathematics, and we shall say that the mathematics that we will",
"Python code for prime number": "Python code for prime number, and also the P8 and C, respectively.\nIn fact, the same rules apply to the P8 range, which is the same in both cases. In either case, we may make a higher number, with the P8 higher; in the case of the P8 lower one,",
"The best way to learn a new language is": "The best way to learn a new language is to read our language, and then get over to the basics. If you're ready to learn a new language, it's easy!\nWhat Do I Need to Know?\nYou are speaking to a native speaker every day, not on the same day, at the same time. That's right"
}
}