diff --git a/src_erl/NerlnetApp/src/MainServer/initHandler.erl b/src_erl/NerlnetApp/src/MainServer/initHandler.erl
index 46af9700..e1a44990 100644
--- a/src_erl/NerlnetApp/src/MainServer/initHandler.erl
+++ b/src_erl/NerlnetApp/src/MainServer/initHandler.erl
@@ -22,8 +22,8 @@ init(Req0, [Main_genServer_Pid]) ->
   %Bindings also can be accessed as once, giving a map of all bindings of Req0:
   {_,Body,_} = cowboy_req:read_body(Req0, #{length => ?DATA_LEN}),  %read up to X MB (default was 8MB)
   Decoded_body = binary_to_list(Body),
-  [Index, TotalSources, SourceName, WorkersStr, NumOfBatches, NerlTensorType, Data] = string:split(Decoded_body, "#", all),
-  gen_server:cast(Main_genServer_Pid,{initCSV, Index, TotalSources, SourceName, WorkersStr, NumOfBatches, NerlTensorType, Data}),
+  [Index, TotalSources, SourceName, WorkersStr, Phase, NumOfBatches, NerlTensorType, Data] = string:split(Decoded_body, "#", all),
+  gen_server:cast(Main_genServer_Pid,{initCSV, Index, TotalSources, SourceName, WorkersStr, Phase, NumOfBatches, NerlTensorType, Data}),
   %[Source|WorkersAndInput] = re:split(binary_to_list(Body), "#", [{return, list}]),
   %{Workers,SourceData} = getWorkerInput(WorkersAndInput,[]),
 
diff --git a/src_erl/NerlnetApp/src/MainServer/mainGenserver.erl b/src_erl/NerlnetApp/src/MainServer/mainGenserver.erl
index 6c61df6e..07659667 100644
--- a/src_erl/NerlnetApp/src/MainServer/mainGenserver.erl
+++ b/src_erl/NerlnetApp/src/MainServer/mainGenserver.erl
@@ -82,13 +82,13 @@ init({MyName,ClientsNames,BatchSize,WorkersMap,NerlnetGraph , DeviceName}) ->
   {ok, #main_genserver_state{myName = MyNameStr , state=idle, total_sources=0, sources_data_ready_ctr = 0}}.
 
 
-handle_cast({initCSV, _Index, TotalSources, SourceName, WorkersList, NumOfBatches, NerlTensorType, Data}, State = #main_genserver_state{state = idle, sourcesWaitingList = SourcesWaitingList, total_sources = TotalSourcesOld, sources_data_ready_ctr = SourcesDataReadyCtrOld}) ->
+handle_cast({initCSV, _Index, TotalSources, SourceName, WorkersList, Phase, NumOfBatches, NerlTensorType, Data}, State = #main_genserver_state{state = idle, sourcesWaitingList = SourcesWaitingList, total_sources = TotalSourcesOld, sources_data_ready_ctr = SourcesDataReadyCtrOld}) ->
   {RouterHost,RouterPort} = ets:lookup_element(get(main_server_ets), my_router, ?DATA_IDX),
   ActionStr = atom_to_list(updateCSV),
   {TotalSourcesInt, _Rest} = string:to_integer(TotalSources),
   % MessageBody = WorkersList ++ "#" ++ NumOfBatches ++ "#" ++ NerlTensorType ++ "#" ++ Data,
   WorkersListSeperated = string:split(WorkersList, ",", all),
-  MessageBody = {WorkersListSeperated, NumOfBatches, NerlTensorType, Data},
+  MessageBody = {WorkersListSeperated, Phase, NumOfBatches, NerlTensorType, Data},
   nerl_tools:http_router_request(RouterHost,RouterPort, [SourceName], ActionStr, MessageBody), % update the source with its data
   UpdatedSourceWaitingList = SourcesWaitingList++[list_to_atom(SourceName)],
   {SourcesDataReadyCtr, NewTotalSources} = 
@@ -98,6 +98,7 @@ handle_cast({initCSV, _Index, TotalSources, SourceName, WorkersList, NumOfBatche
   end,
   {noreply, State#main_genserver_state{sourcesWaitingList = UpdatedSourceWaitingList, total_sources = NewTotalSources, sources_data_ready_ctr = SourcesDataReadyCtr}};
 
+% TODO Guy - I think this pattern is redundant - it is not relevant to the main server state
 handle_cast({initCSV, _Index, _TotalSources, _SourceName ,_SourceData}, State) ->
   ?LOG_ERROR("initCSV is only applicalble when main server is in idle state!"),
   {noreply, State#main_genserver_state{}};
diff --git a/src_erl/NerlnetApp/src/Source/castingHandler.erl b/src_erl/NerlnetApp/src/Source/castingHandler.erl
index 97d62e31..20f83931 100644
--- a/src_erl/NerlnetApp/src/Source/castingHandler.erl
+++ b/src_erl/NerlnetApp/src/Source/castingHandler.erl
@@ -18,8 +18,8 @@ init(Req0, [Action,Source_StateM_Pid]) ->
 %%  io:format("casting handler got Body:~p~n",[Body]),
   case Action of
     csv ->  {_ , Body} = nerl_tools:read_all_data(Req0 , <<>>),
-            {WorkersList, NumOfBatches, NerlTensorType, Data} = binary_to_term(Body),
-            gen_statem:cast(Source_StateM_Pid, {batchList, WorkersList, list_to_integer(NumOfBatches), NerlTensorType , Data});
+            {WorkersList, Phase, NumOfBatches, NerlTensorType, Data} = binary_to_term(Body),
+            gen_statem:cast(Source_StateM_Pid, {batchList, WorkersList, list_to_atom(Phase), list_to_integer(NumOfBatches), NerlTensorType , Data});
     startCasting  ->  {_,Body,_} = cowboy_req:read_body(Req0),
                       gen_statem:cast(Source_StateM_Pid, {startCasting,Body});
     statistics    ->  gen_statem:cast(Source_StateM_Pid, {statistics});
diff --git a/src_erl/NerlnetApp/src/Source/sourceStatem.erl b/src_erl/NerlnetApp/src/Source/sourceStatem.erl
index 4ceb3942..7739244f 100644
--- a/src_erl/NerlnetApp/src/Source/sourceStatem.erl
+++ b/src_erl/NerlnetApp/src/Source/sourceStatem.erl
@@ -25,6 +25,8 @@
 -define(SENDING_FREQUENCY_OVERHEAD_FIX_FACTOR_PERC, 0.75).
 -define(MICRO_TO_MILLI_FACTOR, 0.001).
 
+-define(PHASE_TRAINING_ATOM, training).
+-define(PHASE_PREDICTION_ATOM, prediction).
 
 -record(source_statem_state, {ets_ref, batchesList = [], castingTo=[], myName, source_pid, transmitter_pid = none,csvName="", nerlTensorType}).
 
@@ -77,6 +79,7 @@ init({MyName, WorkersMap, NerlnetGraph, Policy, BatchSize, Frequency , Epochs, T
   ets:insert(EtsRef, {workers_list, []}),
   ets:insert(EtsRef, {csv_name, ""}), % not in use
   ets:insert(EtsRef, {stats_ets, EtsStatsRef}),
+  ets:insert(EtsRef, {current_phase, none}),
   
   {MyRouterHost,MyRouterPort} = nerl_tools:getShortPath(MyName,?MAIN_SERVER_ATOM, NerlnetGraph),
   ets:insert(EtsRef, {my_router,{MyRouterHost,MyRouterPort}}),
@@ -112,7 +115,7 @@ state_name(_EventType, _EventContent, State = #source_statem_state{}) ->
 
 
 %% This cast receive a list of samples to load to the records batchList
-idle(cast, {batchList, WorkersList, NumOfBatches, NerlTensorType, Data}, State) ->
+idle(cast, {batchList, WorkersList, Phase, NumOfBatches, NerlTensorType, Data}, State) ->
   EtsRef = get(source_ets),
   StatsEtsRef = get(source_stats_ets),
   MyName = ets:lookup_element(EtsRef, my_name, ?DATA_IDX),
@@ -120,6 +123,7 @@ idle(cast, {batchList, WorkersList, NumOfBatches, NerlTensorType, Data}, State)
   {NerlTensorBatchesList, SampleSize} = parser:parseCSV(MyName, BatchSize, NerlTensorType, Data), % TODO this is slow and heavy policy! pre parse in ETS a possible solution
   ets:update_element(EtsRef, workers_list, [{?DATA_IDX, WorkersList}]),
   ets:update_element(EtsRef, num_of_batches, [{?DATA_IDX, NumOfBatches}]),
+  ets:update_element(EtsRef, current_phase, [{?DATA_IDX, Phase}]),
   ets:insert(EtsRef, {nerlTensorType, NerlTensorType}),
   stats:increment_messages_received(StatsEtsRef),
   ?LOG_NOTICE("Source ~p, workers are: ~p", [MyName, WorkersList]),
@@ -145,8 +149,8 @@ idle(cast, {startCasting,_Body}, State = #source_statem_state{batchesList = Batc
   BatchSize = ets:lookup_element(EtsRef, batch_size, ?DATA_IDX),
   SampleSize = ets:lookup_element(EtsRef, sample_size, ?DATA_IDX),
   Epochs = ets:lookup_element(EtsRef, epochs, ?DATA_IDX),
-  BatchSize = ets:lookup_element(EtsRef, batch_size, ?DATA_IDX),
   WorkersList = ets:lookup_element(EtsRef, workers_list, ?DATA_IDX),
+  Phase = ets:lookup_element(EtsRef, current_phase, ?DATA_IDX),
 
   %% UserLimitNumberOfBatchesToSendInt = list_to_integer(UserLimitNumberOfBatchesToSend),
   %% BatchesToSend = min(length(BatchesList), UserLimitNumberOfBatchesToSendInt),
@@ -154,7 +158,7 @@ idle(cast, {startCasting,_Body}, State = #source_statem_state{batchesList = Batc
   BatchesListFinal = lists:sublist(BatchesList, BatchesToSend), % Batches list with respect of constraint UserLimitNumberOfBatchesToSendInt
 
   % TODO consider add offset value from API
-  ?LOG_NOTICE("~p - starts casting to workers: ~p",[MyName, WorkersList]),
+  ?LOG_NOTICE("~p - starts casting of phase ~p to workers: ~p",[MyName, Phase, WorkersList]),
   ?LOG_NOTICE("Frequency: ~pHz [Batches/Second]",[Frequency]),
   ?LOG_NOTICE("Batch size: ~p", [BatchSize]),
   ?LOG_NOTICE("Sample size = ~p",[SampleSize]),
@@ -283,6 +287,13 @@ spawnTransmitter(SourceEtsRef, WorkersListOfNames, BatchesListToSend)->
   TimeInterval_ms = ets:lookup_element(SourceEtsRef, time_interval_ms, ?DATA_IDX), % frequency to time interval duration in milliseconds between each send
   ClientWorkerPairs = nerl_tools:get_client_worker_pairs(WorkersListOfNames,WorkersMap,[]),
   Epochs = ets:lookup_element(SourceEtsRef, epochs, ?DATA_IDX),
+  Phase = ets:lookup_element(SourceEtsRef, current_phase, ?DATA_IDX),
+  MyName = ets:lookup_element(SourceEtsRef, my_name, ?DATA_IDX),
+  case Phase of
+    ?PHASE_TRAINING_ATOM -> pass;
+    ?PHASE_PREDICTION_ATOM -> Epochs = 1; % In prediction phase, we send only a single epoch always!
+    _ -> ?LOG_ERROR("Source ~p has an unknown phase: ~p",[MyName, Phase])
+  end,
   SourcePid = self(),
   TimeIntervalWithOverheadFactor = TimeInterval_ms * ?SENDING_FREQUENCY_OVERHEAD_FIX_FACTOR_PERC,
   spawn_link(?MODULE,transmitter,[TimeIntervalWithOverheadFactor,SourceEtsRef, SourcePid ,Epochs, ClientWorkerPairs, BatchesListToSend, Method]).