diff --git a/lib/dictBuilder/cover.c b/lib/dictBuilder/cover.c
index beeff7a508f..30a2c65dee0 100644
--- a/lib/dictBuilder/cover.c
+++ b/lib/dictBuilder/cover.c
@@ -640,25 +640,14 @@ static size_t COVER_ctx_init(COVER_ctx_t *ctx, const void *samplesBuffer,
   const unsigned nbTestSamples = splitPoint < 1.0 ? nbSamples - nbTrainSamples : nbSamples;
   const size_t trainingSamplesSize = splitPoint < 1.0 ? COVER_sum(samplesSizes, nbTrainSamples) : totalSamplesSize;
   const size_t testSamplesSize = splitPoint < 1.0 ? COVER_sum(samplesSizes + nbTrainSamples, nbTestSamples) : totalSamplesSize;
-  const size_t readLength = MAX(d, sizeof(U64);
   ctx->displayLevel = displayLevel;
   /* Checks */
-  if (totalSamplesSize < readLength) ||
+  if (totalSamplesSize < MAX(d, sizeof(U64)) ||
       totalSamplesSize >= (size_t)COVER_MAX_SAMPLES_SIZE) {
     DISPLAYLEVEL(1, "Total samples size is too large (%u MB), maximum size is %u MB\n",
                  (unsigned)(totalSamplesSize>>20), (COVER_MAX_SAMPLES_SIZE >> 20));
     return ERROR(srcSize_wrong);
   }
-  /* The training set must be large enough for the partial suffix array, whose
-   * size is `trainingSamplesSize - MAX(d, sizeof(U64)) + 1`. When splitPoint
-   * < 1.0 the training set is a subset of the corpus, so it can be smaller
-   * than the total even after the check above. Guard against the size_t
-   * subtraction underflowing into a huge allocation. */
-  if (trainingSamplesSize < MAX(d, sizeof(U64))) {
-    DISPLAYLEVEL(1, "Training samples size is too small (%u), minimum size is %u\n",
-                 (unsigned)trainingSamplesSize, (unsigned)MAX(d, sizeof(U64)));
-    return ERROR(srcSize_wrong);
-  }
   /* Check if there are at least 5 training samples */
   if (nbTrainSamples < 5) {
     DISPLAYLEVEL(1, "Total number of training samples is %u and is invalid.", nbTrainSamples);
@@ -681,7 +670,7 @@ static size_t COVER_ctx_init(COVER_ctx_t *ctx, const void *samplesBuffer,
   ctx->nbTrainSamples = nbTrainSamples;
   ctx->nbTestSamples = nbTestSamples;
   /* Partial suffix array */
-  ctx->suffixSize = trainingSamplesSize - readLength + 1;
+  ctx->suffixSize = trainingSamplesSize - MAX(d, sizeof(U64)) + 1;
   ctx->suffix = (U32 *)malloc(ctx->suffixSize * sizeof(U32));
   /* Maps index to the dmerID */
   ctx->dmerAt = (U32 *)malloc(ctx->suffixSize * sizeof(U32));
diff --git a/tests/fuzzer.c b/tests/fuzzer.c
index 4a2f0bac8f8..346ef5c82ef 100644
--- a/tests/fuzzer.c
+++ b/tests/fuzzer.c
@@ -4260,25 +4260,6 @@ static int basicUnitTests(U32 const seed, double compressibility)
         if (dictID==0) goto _output_error;
         DISPLAYLEVEL(3, "OK : %u \n", (unsigned)dictID);
 
-        DISPLAYLEVEL(3, "test%3i : ZDICT_optimizeTrainFromBuffer_cover with tiny training set : ", testNb++);
-        {   /* With splitPoint < 1.0 the training set is a subset of the corpus.
-             * If that subset is smaller than MAX(d, sizeof(U64)) the partial
-             * suffix array size used to underflow size_t (issue #4682). The
-             * builder must reject it cleanly instead. The first few samples are
-             * tiny so the 0.75 training split stays below the threshold, while
-             * the total corpus is comfortably large. */
-            size_t tinySizes[8] = { 1, 1, 1, 1, 1, 1, 100, 100 };
-            ZDICT_cover_params_t tinyParams;
-            size_t tinyResult;
-            memset(&tinyParams, 0, sizeof(tinyParams));
-            tinyParams.steps = 4;
-            tinyParams.splitPoint = 0.75;
-            tinyResult = ZDICT_optimizeTrainFromBuffer_cover(
-                dictBuffer, optDictSize, CNBuffer, tinySizes, 8, &tinyParams);
-            if (!ZDICT_isError(tinyResult)) goto _output_error;
-        }
-        DISPLAYLEVEL(3, "OK \n");
-
         ZSTD_freeCCtx(cctx);
         free(dictBuffer);
         free(samplesSizes);