add rlhf_ranking

Dockerfile

@@ -25,6 +25,8 @@ RUN chmod 777 .
 
 COPY prodigy.json .
 COPY data ./data/
+COPY rlhf-ranking.py .
+COPY prodigy.sh .
 
 ENV PRODIGY_HOME /app
 ENV PRODIGY_LOGGING "verbose"
@@ -32,4 +34,4 @@ ENV PRODIGY_ALLOWED_SESSIONS "user1,user2"
 
 EXPOSE 7860
 
-CMD python -m prodigy ner.manual ner_news en_core_web_sm ./data/dataset.jsonl --label PERSON,ORG,PRODUCT
+CMD CMD ["bash","prodigy.sh"]

prodigy.sh

@@ -0,0 +1 @@
+python -m prodigy rlhf.ranking rlhf_data data/dataset.jsonl -F rlhf_ranking.py

rlhf_ranking.py

@@ -0,0 +1,64 @@
+import prodigy
+import itertools as it
+from prodigy.util import set_hashes
+from prodigy import get_stream
+
+
+@prodigy.recipe(
+    "rlhf.ranking",
+    dataset=("Dataset to save answers to", "positional", None, str),
+    source=("Datafile to load", "positional", None, str),
+)
+def ranking(dataset, source):
+    # Load your own streams from anywhere you want
+    stream = get_stream(source)
+    
+    def prep_stream(stream):
+        for ex in stream:
+            ex['text'] = ex['instruction']
+            del ex['instruction']
+            for c1, c2 in it.combinations(ex['options'], 2):
+                ex['options'] = [
+                    {"id": 0, "text": c1['response']},
+                    {"id": 1, "text": c2['response']}
+                ]
+                yield set_hashes(ex)
+            
+
+    return {
+        "dataset": dataset,
+        "view_id": "choice",
+        "stream": prep_stream(stream),
+        "config":{
+            "global_css": ".prodigy-option{font-size: 15px;}"
+        }
+    }
+
+
+@prodigy.recipe(
+    "rlhf.respond",
+    dataset=("Dataset to save answers to", "positional", None, str),
+    source=("Datafile to load", "positional", None, str),
+)
+def ranking(dataset, source):
+    # Load your own streams from anywhere you want
+    stream = get_stream(source)
+    
+    def prep_stream(stream):
+        for ex in stream:
+            ex['text'] = ex['instruction']
+            del ex['instruction']
+            yield ex
+
+    return {
+        "dataset": dataset,
+        "view_id": "blocks",
+        "stream": prep_stream(stream),
+        "config":{
+            "global_css": ".prodigy-option{font-size: 15px;}",
+            "blocks":[
+                {"view_id": "text"},
+                {"view_id": "text_input", "field_autofocus": True, "field_rows": 4, "field_placeholder": "Try to use 2-3 sentences to answer the question."},
+            ],
+        }
+    }