Replace selected medium Sample 1 with seed 2 sample 125; preserve Sample 2

iclr-debug-medium-speed-full-decay-20260918-v1/decoder-mstok/evaluation/step-26000/selected-samples/medium-selected-samples-step26000.json

@@ -6,19 +6,19 @@
   "tf32_allowed": false,
   "samples": [
     {
-      "seed": 0,
-      "index": 70,
-      "text": " but it is usually easy to do how each that can work for it, that would products things that can be done. Andt means there weren’t tools to help how a logic without implementing any specific code.\n\nNow, on what you’ll have with this, it’s easier to try to solve the problem. There are several examples of new language by programming. For now we are some enforce languages, are new languages, and defined information. You can see the code for the language. comment sounds. Many teach toer usingified is to understand language I understand. However, the idea is that I’ve used in the early processes in writing the data, for you are still for such programs. there are also new languages with some parts that are the other language. For example, it would be wrong to one something from one language language. You can write that with three separate parts of your own is under just many a languages.\n\nIt would require a better language language. This means that in a language in an error is considered critical, i.e. If you write a language or better to make a single and; that we can write a language that provides into very different things. I really like to refine my problems",
+      "seed": 2,
+      "index": 125,
+      "text": " code that works. This is why so does US author provide an example of a concept for the value type.\n\nThere is no explanation as to or given an object which can read the only value or value one of the functions. It is argument to read about whether answer is related to the first case which refers to to and the answer if it is the then certainly related. But it is not clear whether it was a way to put a argument by a object which is written, in order to get written together in order to not remove them.\n\nAnother important thing is that the best type of question is the object. For example, if it is important to remember this is an item in the type of list timei.e., an actual element which requires the, in order to produce the actual scale of the sequence, the same type of item can be done. This is whether it's one or the one queue between the and the part of one, which can be addressed, and is better remembered. In this case there are no questions about this argument: something something are wrong, which means that, at the same end of on same thing, the same type of item is going to use. This is because.'t't find that good that's a simple",
       "prompt": "",
       "token_count": 256,
       "scored_tokens": 255,
-      "gen_ppl": 76.85733032226562,
-      "entropy_nats": 4.504906832692063,
-      "source_file": "/home/ishank/mstok-results/iclr-debug-medium-speed-full-decay-20260918-v1/decoder-mstok/evaluation/step-26000/random-seed-0.samples.json",
-      "text_sha256": "c5f49753a454c2ab3d8bc7f6284be40204d03b032743eb05497128c6da44ddb6",
+      "gen_ppl": 58.0941047668457,
+      "entropy_nats": 4.326089171303893,
+      "source_file": "/home/ishank/mstok-results/iclr-debug-medium-speed-full-decay-20260918-v1/decoder-mstok/evaluation/step-26000/random-seed-2.samples.json",
+      "text_sha256": "1d1171d39e0f6cffc42e938ff7c2180b6589478a5780d7b0805eee097d4fa714",
       "replacement_characters": 0,
-      "adjacent_word_repetition": 0.009345794392523364,
-      "repeated_4gram_fraction": 0
+      "adjacent_word_repetition": 0.008888888888888889,
+      "repeated_4gram_fraction": 0.008968609865470878
     },
     {
       "seed": 4,
@@ -36,6 +36,19 @@
       "repeated_4gram_fraction": 0.004950495049504955
     }
   ],
-  "selection_note": "Curated for topic consistency, absence of long repetition loops, and proximity to requested PPL. Both retain conspicuous grammatical and semantic errors; not fluent or representative random examples.",
-  "score_source": "/home/ishank/mstok-results/medium-sample-scores-step26000-20260919.json"
+  "selection_note": "Curated examples, not representative random draws. Sample 1 was replaced after reviewing 93 exact-256-token candidates without Unicode replacement characters, prioritizing topic continuity and fewer malformed word fragments over the original PPL range. Both retain conspicuous grammatical and semantic errors; not fluent prose.",
+  "score_source": "/home/ishank/mstok-results/medium-sample-scores-step26000-20260919.json",
+  "replacement_history": [
+    {
+      "sample_number": 1,
+      "previous_seed": 0,
+      "previous_index": 70,
+      "previous_gen_ppl": 76.85733032226562,
+      "previous_text_sha256": "c5f49753a454c2ab3d8bc7f6284be40204d03b032743eb05497128c6da44ddb6",
+      "new_seed": 2,
+      "new_index": 125,
+      "previous_hf_commit": "bdcd2eb9e5e3082030239053062d89b282651e76",
+      "sample_2_unchanged": true
+    }
+  ]
 }

iclr-debug-medium-speed-full-decay-20260918-v1/decoder-mstok/evaluation/step-26000/selected-samples/medium-selected-samples-step26000.tex

@@ -2,17 +2,17 @@
 % Supplied true level-zero code: these are NOT unconditional samples.
 % Individual PPL: GPT-2 large, FP32, TF32 disabled.
 \noindent
-% Checkpoint 26000; seed 0; zero-based sample index 70; 256 tokens.
+% Checkpoint 26000; seed 2; zero-based sample index 125; 256 tokens.
 \begin{minipage}[t]{0.49\linewidth}
 \begin{tcolorbox}[samplestyle, colback=white, colframe=black]
 \scriptsize\linespread{1.06}\selectfont
 \spheading{black}{Sample 1}{%
-  \textbf{Gen.\ PPL:} 76.86}
-\noindent but it is usually easy to do how each that can work for it, that would products things that can be done. Andt means there weren’t tools to help how a logic without implementing any specific code.
+  \textbf{Gen.\ PPL:} 58.09}
+\noindent code that works. This is why so does US author provide an example of a concept for the value type.
 
-Now, on what you’ll have with this, it’s easier to try to solve the problem. There are several examples of new language by programming. For now we are some enforce languages, are new languages, and defined information. You can see the code for the language. comment sounds. Many teach toer usingified is to understand language I understand. However, the idea is that I’ve used in the early processes in writing the data, for you are still for such programs. there are also new languages with some parts that are the other language. For example, it would be wrong to one something from one language language. You can write that with three separate parts of your own is under just many a languages.
+There is no explanation as to or given an object which can read the only value or value one of the functions. It is argument to read about whether answer is related to the first case which refers to to and the answer if it is the then certainly related. But it is not clear whether it was a way to put a argument by a object which is written, in order to get written together in order to not remove them.
 
-It would require a better language language. This means that in a language in an error is considered critical, i.e. If you write a language or better to make a single and; that we can write a language that provides into very different things. I really like to refine my problems
+Another important thing is that the best type of question is the object. For example, if it is important to remember this is an item in the type of list timei.e., an actual element which requires the, in order to produce the actual scale of the sequence, the same type of item can be done. This is whether it's one or the one queue between the and the part of one, which can be addressed, and is better remembered. In this case there are no questions about this argument: something something are wrong, which means that, at the same end of on same thing, the same type of item is going to use. This is because.'t't find that good that's a simple
 \end{tcolorbox}
 \end{minipage}\hfill
 % Checkpoint 26000; seed 4; zero-based sample index 55; 256 tokens.
@@ -49,4 +49,3 @@ Once you only face a target at the same time, you can still face it on the groun
 \end{minipage}
 \par\smallskip
 {\scriptsize Selected 256-token outputs from \name-med.\ (step 26,000), conditioned on the true level-zero code. Generated wording is unchanged.}
-