add seed and code to allow multi-sample averaging for variance reduct…

…ion. also run linter.

arxiv_scraper.py

@@ -31,18 +31,24 @@ def __hash__(self):
 
 
 def is_earlier(ts1, ts2):
-    return int(ts1.replace('.','')) < int(ts2.replace('.',''))
+    return int(ts1.replace(".", "")) < int(ts2.replace(".", ""))
 
 
 def get_papers_from_arxiv_api(area: str, timestamp, last_id) -> List[Paper]:
     # grabs papers from the arxiv API endpoint.
     # we need this because the RSS feed is buggy, and drops some of the most recent papers silently.
     end_date = timestamp
-    start_date = timestamp-timedelta(days=4)
+    start_date = timestamp - timedelta(days=4)
     search = arxiv.Search(
-        query="("+area+") AND submittedDate:[" + start_date.strftime('%Y%m%d') + "* TO " + end_date.strftime('%Y%m%d') + "*]",
+        query="("
+        + area
+        + ") AND submittedDate:["
+        + start_date.strftime("%Y%m%d")
+        + "* TO "
+        + end_date.strftime("%Y%m%d")
+        + "*]",
         max_results=None,
-        sort_by=arxiv.SortCriterion.SubmittedDate
+        sort_by=arxiv.SortCriterion.SubmittedDate,
     )
     results = list(arxiv.Client().results(search))
     api_papers = []
@@ -52,12 +58,16 @@ def get_papers_from_arxiv_api(area: str, timestamp, last_id) -> List[Paper]:
             authors = [author.name for author in result.authors]
             summary = result.summary
             summary = unescape(re.sub("\n", " ", summary))
-            paper = Paper(authors=authors, title=result.title, abstract=summary, arxiv_id=result.get_short_id()[:10])
+            paper = Paper(
+                authors=authors,
+                title=result.title,
+                abstract=summary,
+                arxiv_id=result.get_short_id()[:10],
+            )
             api_papers.append(paper)
     return api_papers
 
 
-
 def get_papers_from_arxiv_rss(area: str, config: Optional[dict]) -> List[Paper]:
     # get the feed from http://export.arxiv.org/rss/ and use the updated timestamp to avoid duplicates
     updated = datetime.utcnow() - timedelta(days=1)
@@ -73,7 +83,9 @@ def get_papers_from_arxiv_rss(area: str, config: Optional[dict]) -> List[Paper]:
         return []
     # get the list of entries
     entries = feed.entries
-    timestamp = datetime.strptime(feed.headers["last-modified"], '%a, %d %b %Y %H:%M:%S GMT')
+    timestamp = datetime.strptime(
+        feed.headers["last-modified"], "%a, %d %b %Y %H:%M:%S GMT"
+    )
     # ugly hack: this should be the very oldest paper in the RSS feed that was not put on hold.
     # if ArXiv changes their RSS announcement format this line will break, but we have no other way of getting this info
     last_id = feed.entries[0].id.split("/")[-1]
@@ -117,7 +129,6 @@ def get_papers_from_arxiv_rss_api(area: str, config: Optional[dict]) -> List[Pap
     return merged_paper_list
 
 
-
 if __name__ == "__main__":
     paper_list = get_papers_from_arxiv_rss("math.AC", None)
     print(paper_list)

filter_papers.py

@@ -69,13 +69,18 @@ def call_chatgpt(full_prompt, openai_client, model, num_samples):
         messages=[{"role": "user", "content": full_prompt}],
         temperature=0.0,
         n=int(num_samples),
-        seed=0
+        seed=0,
     )
 
 
 def run_and_parse_chatgpt(full_prompt, openai_client, config):
     # just runs the chatgpt prompt, tries to parse the resulting JSON
-    completion = call_chatgpt(full_prompt, openai_client, config["SELECTION"]["model"], config["FILTERING"]["num_samples"])
+    completion = call_chatgpt(
+        full_prompt,
+        openai_client,
+        config["SELECTION"]["model"],
+        config["FILTERING"]["num_samples"],
+    )
     json_dicts = defaultdict(list)
     for choice in completion.choices:
         out_text = choice.message.content
@@ -99,12 +104,18 @@ def run_and_parse_chatgpt(full_prompt, openai_client, config):
                 continue
     all_dict = []
     for id, json_list in json_dicts.items():
-        rel_score = sum([float(jdict["RELEVANCE"]) for jdict in json_list])/float(len(json_list))
-        nov_score = sum([float(jdict["NOVELTY"]) for jdict in json_list]) / float(len(json_list))
-        new_dict = {"ARXIVID":json_list[0]["ARXIVID"],
-        "COMMENT":json_list[0]["COMMENT"],
-        "RELEVANCE":rel_score,
-        "NOVELTY":nov_score}
+        rel_score = sum([float(jdict["RELEVANCE"]) for jdict in json_list]) / float(
+            len(json_list)
+        )
+        nov_score = sum([float(jdict["NOVELTY"]) for jdict in json_list]) / float(
+            len(json_list)
+        )
+        new_dict = {
+            "ARXIVID": json_list[0]["ARXIVID"],
+            "COMMENT": json_list[0]["COMMENT"],
+            "RELEVANCE": rel_score,
+            "NOVELTY": nov_score,
+        }
         all_dict.append(new_dict)
     return all_dict, calc_price(config["SELECTION"]["model"], completion.usage)
 

parse_json_to_md.py

@@ -67,7 +67,7 @@ def render_md_string(papers_dict):
     # join all papers into one string
     output_string = output_string + "\n".join(paper_strings)
     output_string += "\n\n---\n\n"
-    output_string += f'## Paper selection prompt\n{criterion}'
+    output_string += f"## Paper selection prompt\n{criterion}"
     return output_string