diff --git a/tokenizers/src/models/bpe/trainer.rs b/tokenizers/src/models/bpe/trainer.rs
index 2876f1ef5..81468ad2f 100644
--- a/tokenizers/src/models/bpe/trainer.rs
+++ b/tokenizers/src/models/bpe/trainer.rs
@@ -8,6 +8,63 @@ use serde::{Deserialize, Serialize};
 use std::cmp::Ordering;
 use std::collections::{BinaryHeap, HashMap, HashSet};
 
+mod unsafevec {
+    use std::cell::UnsafeCell;
+    use std::marker::PhantomData;
+
+    use super::Word;
+
+    pub(super) struct UnsafeVec<'v> {
+        value: UnsafeCell<*mut Word>,
+        len: usize,
+        _data: PhantomData<&'v mut Vec<Word>>,
+    }
+
+    // SAFETY: the only use of a &UnsafeVec is to pass it to
+    // fn get_unchecked_mut(), that function and its safety contract
+    // mandate that only disjoint indexes may be used
+    // if that unsafe contract is upheld there's no way to abuse a &UnsafeVec
+    // to obtain unsynchronized data access
+    unsafe impl<'v> Sync for UnsafeVec<'v> {}
+
+    impl<'v> UnsafeVec<'v> {
+        pub(super) fn new(words: &'v mut Vec<Word>) -> UnsafeVec<'v> {
+            let len = words.len();
+            UnsafeVec {
+                value: UnsafeCell::new(words.as_mut_ptr()),
+                len,
+                _data: PhantomData,
+            }
+        }
+        // SAFETY: per UnsafeVec, every call to get_unchecked_mut must be made with a unique index i
+        #[allow(clippy::mut_from_ref)]
+        pub(super) unsafe fn get_unchecked_mut(&self, i: isize) -> &mut Word {
+            assert!((i as usize) < self.len);
+            &mut *(*self.value.get()).offset(i)
+        }
+    }
+    // run with miri: cargo +nightly miri test test_unsafe_vec
+    #[test]
+    fn test_unsafe_vec() {
+        let mut v: Vec<_> = std::iter::repeat(Word::new()).take(100).collect();
+        let vec = UnsafeVec::new(&mut v);
+        // if these ranges overlap this is immediately detected by miri
+        let [first, second] = [0..50, 50..100];
+        std::thread::scope(|s| {
+            s.spawn(|| unsafe {
+                for i in first {
+                    *vec.get_unchecked_mut(i) = Word::new();
+                }
+            });
+            s.spawn(|| unsafe {
+                for i in second {
+                    *vec.get_unchecked_mut(i) = Word::new();
+                }
+            });
+        });
+    }
+}
+
 #[derive(Debug, Eq)]
 struct Merge {
     pair: Pair,
@@ -454,7 +511,7 @@ impl BpeTrainer {
         // 3. Tokenize words
         //
         self.update_progress(&progress, word_counts.len(), "Tokenize words");
-        let (words, counts) =
+        let (mut words, counts) =
             self.tokenize_words(word_counts, &mut word_to_id, &mut id_to_word, &progress);
         self.finalize_progress(&progress, words.len());
 
@@ -529,24 +586,27 @@ impl BpeTrainer {
             }
             merges.push((top.pair, new_token_id));
 
-            // Merge the new pair in every words
-            let changes = top
-                .pos
-                .maybe_par_iter()
-                .flat_map(|&i| {
-                    let word = &words[i] as *const _ as *mut Word;
-                    // We can merge each of these words in parallel here because each position
-                    // can be there only once (HashSet). So this is safe.
-                    unsafe {
-                        // let word: &mut Word = &mut (*word);
-                        (*word)
-                            .merge(top.pair.0, top.pair.1, new_token_id, max_token_length)
-                            .into_iter()
-                            .map(|c| (c, i))
-                            .collect::<Vec<_>>()
-                    }
-                })
-                .collect::<Vec<_>>();
+            let changes = {
+                let words = unsafevec::UnsafeVec::new(&mut words);
+
+                // Merge the new pair in every words
+                top.pos
+                    .maybe_par_iter()
+                    .flat_map(|&i| {
+                        // We can merge each of these words in parallel here because each position
+                        // can be there only once (HashSet). So this is safe.
+                        // SAFETY: this really is a HashSet, making i's unique
+                        let _: &HashSet<_> = &top.pos;
+                        unsafe {
+                            let word: &mut Word = words.get_unchecked_mut(i as isize);
+                            word.merge(top.pair.0, top.pair.1, new_token_id, max_token_length)
+                                .into_iter()
+                                .map(|c| (c, i))
+                                .collect::<Vec<_>>()
+                        }
+                    })
+                    .collect::<Vec<_>>()
+            };
 
             // Introduce new formed pairs
             for ((pair, change), iw) in changes {
diff --git a/tokenizers/src/utils/parallelism.rs b/tokenizers/src/utils/parallelism.rs
index b955731d1..903f3b533 100644
--- a/tokenizers/src/utils/parallelism.rs
+++ b/tokenizers/src/utils/parallelism.rs
@@ -32,7 +32,11 @@ pub fn get_parallelism() -> bool {
             v.make_ascii_lowercase();
             !matches!(v.as_ref(), "" | "off" | "false" | "f" | "no" | "n" | "0")
         }
+        #[cfg(not(miri))]
         Err(_) => true, // If we couldn't get the variable, we use the default
+        // FIXME: for now turn parallelism off under miri, otherwise complains about crossbeam-epoch
+        #[cfg(miri)]
+        Err(_) => false,
     }
 }