| { |
| "model_id": "m08", |
| "checkpoint": "results/m08/final.pt", |
| "trained_step": 15258, |
| "total_params": 58774528, |
| "active_params": 58774528, |
| "flops_per_forward": 117549056.0, |
| "val_loss": 3.468865367770195, |
| "val_perplexity": 32.10029973945804, |
| "bits_per_token": 5.00451486359353, |
| "ppl_tokens_evaluated": 655360, |
| "prefill_tok_per_sec": 176358.66558708608, |
| "inference_tok_per_sec": 215.73418372595413, |
| "inference_tok_per_sec_kvcache": 210.01077375697125, |
| "kvcache_speedup": 0.9734700830895984, |
| "gen_tokens": 128, |
| "peak_infer_vram_gb": 0.366, |
| "peak_train_vram_gb": 3.993, |
| "closeness_answer_ppl": 33.412, |
| "closeness_hit_rate": 0.35, |
| "closeness_token_f1": 0.017, |
| "closeness_n": 20, |
| "closeness_items": [ |
| { |
| "prompt": "The capital of France is", |
| "answer": " Paris", |
| "generated": " the capital of the French Republic", |
| "answer_nll": 3.044, |
| "answer_ppl": 20.99, |
| "hit": false, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "The capital of Japan is", |
| "answer": " Tokyo", |
| "generated": " the capital of the Republic of", |
| "answer_nll": 3.2951, |
| "answer_ppl": 26.979, |
| "hit": false, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "The largest planet in the solar system is", |
| "answer": " Jupiter", |
| "generated": " the solar system.\nThe", |
| "answer_nll": 4.1156, |
| "answer_ppl": 61.291, |
| "hit": false, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "Water is made of hydrogen and", |
| "answer": " oxygen", |
| "generated": " oxygen.\nThe hydrogen atom", |
| "answer_nll": 1.0477, |
| "answer_ppl": 2.851, |
| "hit": true, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "The opposite of hot is", |
| "answer": " cold", |
| "generated": " the same.\nThe first", |
| "answer_nll": 3.7453, |
| "answer_ppl": 42.323, |
| "hit": false, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "The opposite of up is", |
| "answer": " down", |
| "generated": " the fact that the two are", |
| "answer_nll": 8.439, |
| "answer_ppl": 4623.824, |
| "hit": false, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "Two plus two equals", |
| "answer": " four", |
| "generated": " two equal numbers of the two", |
| "answer_nll": 2.7378, |
| "answer_ppl": 15.452, |
| "hit": false, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "The sun rises in the", |
| "answer": " east", |
| "generated": " sky, and the sun is", |
| "answer_nll": 4.242, |
| "answer_ppl": 69.545, |
| "hit": false, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "The first president of the United States was George", |
| "answer": " Washington", |
| "generated": " Washington. He was the first", |
| "answer_nll": 0.3207, |
| "answer_ppl": 1.378, |
| "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": 0.0622, |
| "answer_ppl": 1.064, |
| "hit": true, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "A triangle has three", |
| "answer": " sides", |
| "generated": " sides, and the triangle has", |
| "answer_nll": 1.5957, |
| "answer_ppl": 4.932, |
| "hit": true, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "The chemical symbol for gold is", |
| "answer": " Au", |
| "generated": " the gold.\nThe symbol", |
| "answer_nll": 8.9649, |
| "answer_ppl": 7823.719, |
| "hit": false, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "The freezing point of water in Celsius is", |
| "answer": " zero", |
| "generated": " about 1.5\u00b0C", |
| "answer_nll": 5.2632, |
| "answer_ppl": 193.092, |
| "hit": false, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "The author of Romeo and Juliet is William", |
| "answer": " Shakespeare", |
| "generated": " Shakespeare, the author of Romeo", |
| "answer_nll": 0.4581, |
| "answer_ppl": 1.581, |
| "hit": true, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "The Earth revolves around the", |
| "answer": " Sun", |
| "generated": " sun, and the sun is", |
| "answer_nll": 1.2541, |
| "answer_ppl": 3.505, |
| "hit": true, |
| "token_f1": 0.333 |
| }, |
| { |
| "prompt": "The plural of mouse is", |
| "answer": " mice", |
| "generated": " a word that is used to", |
| "answer_nll": 8.9012, |
| "answer_ppl": 7340.772, |
| "hit": false, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "Photosynthesis occurs in the", |
| "answer": " leaves", |
| "generated": " soil, and it is a", |
| "answer_nll": 4.1607, |
| "answer_ppl": 64.115, |
| "hit": false, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "The human heart pumps", |
| "answer": " blood", |
| "generated": " through the heart.\nThe", |
| "answer_nll": 3.1444, |
| "answer_ppl": 23.205, |
| "hit": false, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "The currency used in the United States is the", |
| "answer": " dollar", |
| "generated": " US dollar.\nThe US", |
| "answer_nll": 2.9611, |
| "answer_ppl": 19.32, |
| "hit": true, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "The tallest animal in the world is the", |
| "answer": " giraffe", |
| "generated": " giant tortoise.\nThe", |
| "answer_nll": 2.4257, |
| "answer_ppl": 11.311, |
| "hit": false, |
| "token_f1": 0.0 |
| } |
| ], |
| "generation_samples": { |
| "The most important thing about sleep is": "The most important thing about sleep is that it\u2019s easier to sleep when you\u2019re hungry than when you\u2019re thirsty.\nIt\u2019s important to focus on the most basic sleep patterns you can expect. If you\u2019re hungry, you\u2019ve got the best night\u2019s rest,", |
| "Scientists have discovered that": "Scientists have discovered that a large asteroid that\u2019s about the size of a football.\nThe asteroid hit the Earth in 1999. The asteroid is about the size of Alaska that\u2019s hit the globe in 1999.\nBut asteroids also hit our planet, and it\u2019s likely that the asteroid has killed", |
| "In the history of mathematics,": "In the history of mathematics, the ancient Greek philosopher and physician, Theologus\u2014the Greek philosopher and physician\u2014is described as a human being. The term \u201cGreek\u201d is widely used in the 19th Century, but the word \u201cGreek\u201d is a derivative of the Greek word \u201cepicos", |
| "Python code for prime number": "Python code for prime number systems, for example. However, the main difference between two solutions is their size. In the case where two solutions are identical, the first solution is called the highest number of prime numbers, so the second solution is called the worst number of prime numbers.\nThe numbers in the prime number list,", |
| "The best way to learn a new language is": "The best way to learn a new language is to learn about it.\nThe first step in learning new language is to learn something new.\nThe first step to learning new language is to learn new words.\n\"Reading is not the only thing that can teach us anything. We have to learn anything we can learn. You can learn about" |
| } |
| } |
