CRFEntityExtractor fails to retrieve simple entities from German

I have a small dataset for training, with some very simple entities to retrieve. For instance:

## intent:MY_INTENT
[ich](pronomen) [möchte](fordern)  [eine](menge) schwarzer kaffee.
[ich](pronomen) [möchte](fordern)  diese dinge.

Rasa detects all the intents entirely correctly. But it fails to recognize single entities. For instance if you request ich the entities response will be empty. The same with möchte:

{
  "intent": {
    "name": "MY_INTENT",
    "confidence": 0.9752551317214966
  },
  "entities": [],
  "intent_ranking": [
    {
      "name": "MY_INTENT",
      "confidence": 0.9752551317214966
    },
  ],
  "text": "ich möchte"
}

The weirdest thing is that if I feed there the whole sentence, for instance ich möchte eine kaffee then Rasa captures all or most of the entities. What am I doing wrong?

ich möchte eine personalmutation  
{
  "intent": {
    "name": "MY_INTENT",
    "confidence": 0.999994158744812
  },
  "entities": [
    {
      "start": 0,
      "end": 3,
      "value": "ich",
      "entity": "pronomen",
      "confidence": 0.8411681473886646,
      "extractor": "CRFEntityExtractor"
    },
    {
      "start": 4,
      "end": 10,
      "value": "möchte",
      "entity": "fordern",
      "confidence": 0.8246471681888136,
      "extractor": "CRFEntityExtractor"
    },
    {
      "start": 11,
      "end": 15,
      "value": "eine",
      "entity": "menge",
      "confidence": 0.832780777509731,
      "extractor": "CRFEntityExtractor"
    }
  ],

  "text": "ich möchte eine kaffee"
}


Get more training examples.

btw it’s “Ich moechte einen schwarzen Kaffe” in german