diff --git a/src/djcdata/interface/trainData.h b/src/djcdata/interface/trainData.h
index 3ef9827..b5d81d6 100644
--- a/src/djcdata/interface/trainData.h
+++ b/src/djcdata/interface/trainData.h
@@ -229,7 +229,10 @@ class trainData{
     void readMetaDataFromFile(const std::string& filename);
 
     std::vector<int64_t> getFirstRowsplits()const;
-    std::vector<int64_t> readShapesAndRowSplitsFromFile(const std::string& filename, bool checkConsistency=true);
+    // Returns all distinct rowsplit vectors present in this trainData (one per ragged array group).
+    std::vector<std::vector<int64_t>> getAllRowsplits()const;
+    // Returns all distinct rowsplit vectors found in the file (one per ragged array, deduplicated).
+    std::vector<std::vector<int64_t>> readShapesAndRowSplitsFromFile(const std::string& filename);
 
     void clear();
 
@@ -315,7 +318,8 @@ class trainData{
     void checkFile(FILE *& f, const std::string& filename="")const;
 
 
-    void readRowSplitArray(FILE *&, std::vector<int64_t> &rs, bool check)const;
+    // Reads one typeContainer's worth of arrays from the file, collecting all distinct rowsplits found.
+    void readRowSplitArray(FILE *&, std::vector<std::vector<int64_t>> &all_rs)const;
 
     std::vector<std::vector<int> > getShapes(const typeContainer& a)const;
 
diff --git a/src/djcdata/interface/trainDataGenerator.h b/src/djcdata/interface/trainDataGenerator.h
index 3a2130f..d3e55d6 100644
--- a/src/djcdata/interface/trainDataGenerator.h
+++ b/src/djcdata/interface/trainDataGenerator.h
@@ -127,7 +127,8 @@ class trainDataGenerator{
     std::vector<std::string> orig_infiles_;
     std::vector<size_t> shuffle_indices_;
     std::vector<std::vector<size_t> > sub_shuffle_indices_;
-    std::vector<std::vector<int64_t> > orig_rowsplits_;
+    // outer index: file index; inner vector: all distinct rowsplits for that file (one per ragged array group)
+    std::vector<std::vector<std::vector<int64_t> > > orig_rowsplits_;
     std::vector<size_t> splits_;
     std::vector<bool> usebatch_;
     int randomcount_;
diff --git a/src/djcdata/src/trainData.cpp b/src/djcdata/src/trainData.cpp
index 0533e6d..3c34e3e 100644
--- a/src/djcdata/src/trainData.cpp
+++ b/src/djcdata/src/trainData.cpp
@@ -439,36 +439,39 @@ std::vector<int64_t> trainData::getFirstRowsplits()const{
     return std::vector<int64_t>();
 }
 
-std::vector<int64_t> trainData::readShapesAndRowSplitsFromFile(const std::string& filename, bool checkConsistency){
-    std::vector<int64_t> rowsplits;
+std::vector<std::vector<int64_t>> trainData::getAllRowsplits()const{
+    std::vector<std::vector<int64_t>> out;
+    const std::vector<const typeContainer* > vv = {&feature_arrays_, &truth_arrays_, &weight_arrays_};
+    for(const auto& a: vv){
+        for(size_t i=0;i<a->size();i++){
+            const auto& rs = a->at(i).rowsplits();
+            if(!rs.size()) continue;
+            bool found = false;
+            for(const auto& existing: out)
+                if(existing == rs){ found=true; break; }
+            if(!found)
+                out.push_back(rs);
+        }
+    }
+    return out;
+}
+
+std::vector<std::vector<int64_t>> trainData::readShapesAndRowSplitsFromFile(const std::string& filename){
+    std::vector<std::vector<int64_t>> all_rowsplits;
 
     FILE *ifile = fopen(filename.data(), "rb");
     checkFile(ifile,filename);
 
-    //shapes
-    std::vector<std::vector<int> > dummy;
     readNested(feature_shapes_, ifile);
     readNested(truth_shapes_, ifile);
     readNested(weight_shapes_, ifile);
 
-    //features
-    readRowSplitArray(ifile,rowsplits,checkConsistency);
-    if(!checkConsistency && rowsplits.size()){
-        fclose(ifile);
-        return rowsplits;
-    }
-    //truth
-    readRowSplitArray(ifile,rowsplits,checkConsistency);
-    if(!checkConsistency && rowsplits.size()){
-        fclose(ifile);
-        return rowsplits;
-    }
-    //weights
-    readRowSplitArray(ifile,rowsplits,checkConsistency);
+    readRowSplitArray(ifile, all_rowsplits); // features
+    readRowSplitArray(ifile, all_rowsplits); // truth
+    readRowSplitArray(ifile, all_rowsplits); // weights
 
     fclose(ifile);
-    return rowsplits;
-
+    return all_rowsplits;
 }
 
 void trainData::clear() {
@@ -488,17 +491,17 @@ void trainData::checkFile(FILE *& ifile, const std::string& filename)const{
 
 }
 
-void trainData::readRowSplitArray(FILE *& ifile, std::vector<int64_t> &rowsplits, bool check)const{
+void trainData::readRowSplitArray(FILE *& ifile, std::vector<std::vector<int64_t>> &all_rowsplits)const{
     size_t size = 0;
     io::readFromFile(&size, ifile);
     for(size_t i=0;i<size;i++){
         auto frs = simpleArrayBase::readRowSplitsFromFileP(ifile, true);
         if(frs.size()){
-            if(check){
-                if(rowsplits.size() && rowsplits != frs)
-                    throw std::runtime_error("trainData::readShapesAndRowSplitsFromFile: row splits inconsistent");
-            }
-            rowsplits=frs;
+            bool found = false;
+            for(const auto& existing: all_rowsplits)
+                if(existing == frs){ found=true; break; }
+            if(!found)
+                all_rowsplits.push_back(frs);
         }
     }
 }
diff --git a/src/djcdata/src/trainDataGenerator.cpp b/src/djcdata/src/trainDataGenerator.cpp
index 57642a6..14788d3 100644
--- a/src/djcdata/src/trainDataGenerator.cpp
+++ b/src/djcdata/src/trainDataGenerator.cpp
@@ -40,11 +40,10 @@ void trainDataGenerator::setBuffer(const trainData& td){
     clear();
     if(td.featureShapes().size()<1 || td.featureShapes().at(0).size()<1)
         throw std::runtime_error("trainDataGenerator<T>::setBuffer: no features filled in trainData object");
-    auto hasRagged = tdHasRaggedDimension(td);
 
-    auto rs = td.getFirstRowsplits();
-    if(rs.size())
-        orig_rowsplits_.push_back(rs);
+    auto allrs = td.getAllRowsplits();
+    if(allrs.size())
+        orig_rowsplits_.push_back(allrs);
     shuffle_indices_.push_back(0);
     std::vector<size_t> vec;
     for(size_t i=0;i<td.nElements();i++)
@@ -113,10 +112,11 @@ void trainDataGenerator::readInfo(){
             hasRagged = tdHasRaggedDimension(td);
         }
         if(hasRagged){
-            std::vector<int64_t> rowsplits = td.readShapesAndRowSplitsFromFile(f, firstfile);//check consistency only for first
+            auto allrs = td.readShapesAndRowSplitsFromFile(f);
             if(debuglevel>1)
-                std::cout << "rowsplits.size() " <<rowsplits.size() << ": "<<f <<  std::endl; //debuglevel
-            orig_rowsplits_.push_back(rowsplits);
+                std::cout << "rowsplits groups: " << allrs.size() << " (first size: " << (allrs.size()?allrs.at(0).size():0) << "): " << f << std::endl;
+            if(allrs.size())
+                orig_rowsplits_.push_back(allrs);
         }
         firstfile=false;
         ntotal_ += td.nElements();
@@ -172,13 +172,30 @@ void trainDataGenerator::prepareSplitting(){
     std::vector<int64_t> allrs;
     for(size_t i=0;i<orig_rowsplits_.size();i++){
         auto shuffled_idx = shuffle_indices_.at(i);
-        auto thisrs = orig_rowsplits_.at(shuffled_idx); //inject by file shuffle here
-        thisrs = subShuffleRowSplits(thisrs, sub_shuffle_indices_.at(shuffled_idx));
+        const auto& file_rss = orig_rowsplits_.at(shuffled_idx); //inject by file shuffle here
+
+        // Build a combined rowsplit for this file by summing sub-elements per event across all rowsplit groups.
+        // All groups must have the same number of events (same rowsplit length).
+        std::vector<int64_t> file_combined_rs;
+        for(const auto& rs: file_rss){
+            auto shuffled_rs = subShuffleRowSplits(rs, sub_shuffle_indices_.at(shuffled_idx));
+            if(file_combined_rs.empty()){
+                file_combined_rs = shuffled_rs;
+            } else {
+                auto nelems_a = simpleArrayBase::dataSplitToSplitIndices(file_combined_rs);
+                auto nelems_b = simpleArrayBase::dataSplitToSplitIndices(shuffled_rs);
+                if(nelems_a.size() != nelems_b.size())
+                    throw std::runtime_error("trainDataGenerator::prepareSplitting: rowsplit groups have different numbers of events in the same file");
+                for(size_t j=0;j<nelems_a.size();j++)
+                    nelems_a.at(j) += nelems_b.at(j);
+                file_combined_rs = simpleArrayBase::splitToDataSplitIndices(nelems_a);
+            }
+        }
 
         if(i==0 || allrs.size()==0){
-            allrs=thisrs;}
-        else{
-            allrs = simpleArrayBase::mergeRowSplits(allrs,thisrs);
+            allrs = file_combined_rs;
+        } else {
+            allrs = simpleArrayBase::mergeRowSplits(allrs, file_combined_rs);
         }
     }