Here we first pick the top-k tokens from the distribution of logits, and then sample from them.
\nHere's an experiment that uses these sampling techniques.
\n": "\u3053\u3053\u3067\u306f\u3001\u6700\u521d\u306b\u30ed\u30b8\u30c3\u30c8\u306e\u5206\u5e03\u304b\u3089\u4e0a\u4f4dk\u500b\u306e\u30c8\u30fc\u30af\u30f3\u3092\u9078\u629e\u3057\u3001\u6b21\u306b\u305d\u308c\u3089\u304b\u3089\u30b5\u30f3\u30d7\u30ea\u30f3\u30b0\u3057\u307e\u3059\u3002
\n\u3053\u308c\u306f\u3001\u3053\u308c\u3089\u306e\u30b5\u30f3\u30d7\u30ea\u30f3\u30b0\u624b\u6cd5\u3092\u4f7f\u7528\u3057\u305f\u5b9f\u9a13\u3067\u3059\u3002
\n", "Sample from logits
\n": "\u30ed\u30b8\u30c3\u30c8\u304b\u3089\u306e\u30b5\u30f3\u30d7\u30eb
\n", "New logits filled with _^_0_^_; i.e. zero probability
\n": "_^_0_^_\u65b0\u3057\u3044\u30ed\u30b8\u30c3\u30c8\u3092\u57cb\u3081\u308b\u3001\u3064\u307e\u308a\u78ba\u7387\u304c\u30bc\u30ed
\n", "Pick the largest _^_0_^_ logits and their indices
\n": "_^_0_^_\u6700\u5927\u306e\u30ed\u30b8\u30c3\u30c8\u3068\u305d\u306e\u30a4\u30f3\u30c7\u30c3\u30af\u30b9\u3092\u9078\u629e\u3057\u3066\u304f\u3060\u3055\u3044
\n", "Sample from the top-k logits with the specified sampler.
\n": "\u6307\u5b9a\u3055\u308c\u305f\u30b5\u30f3\u30d7\u30e9\u30fc\u3092\u4f7f\u7528\u3057\u3066\u3001\u4e0a\u304b\u3089k\u500b\u306e\u30ed\u30b8\u30c3\u30c8\u3092\u30b5\u30f3\u30d7\u30ea\u30f3\u30b0\u3057\u307e\u3059\u3002
\n", "Set the values of the top-k selected indices to actual logits. Logits of other tokens remain _^_0_^_
\n": "\u9078\u629e\u3057\u305f\u4e0a\u4f4dk\u306e\u30a4\u30f3\u30c7\u30c3\u30af\u30b9\u306e\u5024\u3092\u5b9f\u969b\u306e\u30ed\u30b8\u30c3\u30c8\u306b\u8a2d\u5b9a\u3057\u307e\u3059\u3002\u4ed6\u306e\u30c8\u30fc\u30af\u30f3\u306e\u30ed\u30b8\u30c3\u30c8\u306f\u6b8b\u308a\u307e\u3059 _^_0_^_
\n", "_^_2_^_ can be any sampler that takes a logits tensor as input and returns a token tensor; e.g. `TemperatureSampler'.
\n": "