| { |
| "model_id": "m28", |
| "checkpoint": "/home/akshat/codemindv2/results/m28/final.pt", |
| "trained_step": 15258, |
| "total_params": 58774528, |
| "active_params": 58774528, |
| "flops_per_forward": 117549056.0, |
| "val_loss": 3.4364620834589004, |
| "val_perplexity": 31.076816264456788, |
| "bits_per_token": 4.9577668060091105, |
| "ppl_tokens_evaluated": 655360, |
| "prefill_tok_per_sec": 191174.46259175235, |
| "inference_tok_per_sec": 215.11173153691215, |
| "inference_tok_per_sec_kvcache": 205.6556695510074, |
| "kvcache_speedup": 0.9560411609429952, |
| "gen_tokens": 128, |
| "peak_infer_vram_gb": 0.366, |
| "peak_train_vram_gb": 3.993, |
| "closeness_answer_ppl": 38.273, |
| "closeness_hit_rate": 0.35, |
| "closeness_token_f1": 0.02, |
| "closeness_n": 20, |
| "closeness_items": [ |
| { |
| "prompt": "The capital of France is", |
| "answer": " Paris", |
| "generated": " the city of Paris. The", |
| "answer_nll": 2.9576, |
| "answer_ppl": 19.252, |
| "hit": true, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "The capital of Japan is", |
| "answer": " Tokyo", |
| "generated": " located in the heart of the", |
| "answer_nll": 4.5123, |
| "answer_ppl": 91.132, |
| "hit": false, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "The largest planet in the solar system is", |
| "answer": " Jupiter", |
| "generated": " the sun.\nThe sun", |
| "answer_nll": 3.5618, |
| "answer_ppl": 35.228, |
| "hit": false, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "Water is made of hydrogen and", |
| "answer": " oxygen", |
| "generated": " helium.\nThe hydrogen is", |
| "answer_nll": 1.0301, |
| "answer_ppl": 2.801, |
| "hit": false, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "The opposite of hot is", |
| "answer": " cold", |
| "generated": " the heat of the sun.", |
| "answer_nll": 5.4039, |
| "answer_ppl": 222.282, |
| "hit": false, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "The opposite of up is", |
| "answer": " down", |
| "generated": " the fact that the two of", |
| "answer_nll": 8.4353, |
| "answer_ppl": 4606.82, |
| "hit": false, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "Two plus two equals", |
| "answer": " four", |
| "generated": " 1.\n- The first", |
| "answer_nll": 4.2668, |
| "answer_ppl": 71.295, |
| "hit": false, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "The sun rises in the", |
| "answer": " east", |
| "generated": " sky, and the sun is", |
| "answer_nll": 3.1628, |
| "answer_ppl": 23.637, |
| "hit": false, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "The first president of the United States was George", |
| "answer": " Washington", |
| "generated": " Washington. He was a man", |
| "answer_nll": 0.8406, |
| "answer_ppl": 2.318, |
| "hit": true, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "The speed of light is approximately 300,000 kilometers per", |
| "answer": " second", |
| "generated": " second. The speed of light", |
| "answer_nll": 0.229, |
| "answer_ppl": 1.257, |
| "hit": true, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "A triangle has three", |
| "answer": " sides", |
| "generated": " sides, and the other two", |
| "answer_nll": 1.4465, |
| "answer_ppl": 4.248, |
| "hit": true, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "The chemical symbol for gold is", |
| "answer": " Au", |
| "generated": " the same as the symbol for", |
| "answer_nll": 8.2115, |
| "answer_ppl": 3683.169, |
| "hit": false, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "The freezing point of water in Celsius is", |
| "answer": " zero", |
| "generated": " the temperature at which the water", |
| "answer_nll": 5.6623, |
| "answer_ppl": 287.822, |
| "hit": false, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "The author of Romeo and Juliet is William", |
| "answer": " Shakespeare", |
| "generated": " Shakespeare.\nThe author of", |
| "answer_nll": 0.5891, |
| "answer_ppl": 1.802, |
| "hit": true, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "The Earth revolves around the", |
| "answer": " Sun", |
| "generated": " sun.\nThe sun is", |
| "answer_nll": 1.6502, |
| "answer_ppl": 5.208, |
| "hit": true, |
| "token_f1": 0.4 |
| }, |
| { |
| "prompt": "The plural of mouse is", |
| "answer": " mice", |
| "generated": " a mouse, which is a", |
| "answer_nll": 5.9685, |
| "answer_ppl": 390.937, |
| "hit": false, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "Photosynthesis occurs in the", |
| "answer": " leaves", |
| "generated": " body.\nThe body is", |
| "answer_nll": 4.1835, |
| "answer_ppl": 65.594, |
| "hit": false, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "The human heart pumps", |
| "answer": " blood", |
| "generated": " are not the only ones that", |
| "answer_nll": 4.1331, |
| "answer_ppl": 62.371, |
| "hit": false, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "The currency used in the United States is the", |
| "answer": " dollar", |
| "generated": " U.S. dollar.", |
| "answer_nll": 3.7576, |
| "answer_ppl": 42.845, |
| "hit": true, |
| "token_f1": 0.0 |
| }, |
| { |
| "prompt": "The tallest animal in the world is the", |
| "answer": " giraffe", |
| "generated": " bull-like bull-like", |
| "answer_nll": 2.8921, |
| "answer_ppl": 18.031, |
| "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 a good time to prepare yourself and maintain a calm and relaxed lifestyle. If we all enjoy the sun, we don\u2019t want to be alone. Our daily routine is a reflection of ourselves, and also our physical health.\nMost of us have probably heard of \ufffd", |
| "Scientists have discovered that": "Scientists have discovered that, in many ways, the human body is actually a computer system. \"We've discovered that when you put a computer in a box, you're going to turn the human body into a computer,\" says Dr. John O'Brien, one of the authors of this piece. \"You can't", |
| "In the history of mathematics,": "In the history of mathematics, the term has been used to describe the practice of using the metric system to measure the relationship between the two measures of the metric system. The metric system, on the other hand, refers to the measurement data of the meter which can be combined into a metric system. The metric system was developed in the", |
| "Python code for prime number": "Python code for prime number of objects.\nEach step in the program is numbered like a number, and a symbol. The function will use the number of the first digits to represent the second digit of the sequence of numbers. The result is the following:\nThis is the function of\n- the first digit of the sequence", |
| "The best way to learn a new language is": "The best way to learn a new language is to learn about it. So if you have an understanding of a language, you will benefit from reading a list of words to learn a new language.\n- Linguistic studies help students learn what the language in which they speak is. Linguistic methods are the study of language with a particular" |
| } |
| } |
