Add 6 more tasks (#78)

- add 6 new tasks (3967c9ef3ac7c0d20a8581fb6ac24032ae6f9f24)
- add print df (6f030e8c53fd09b7034ab155219f8c854e326ed2)
- revert logic (bd9a9adaf078dbd8df2a3532d30ffa2ebbf3ebee)
- fix comment (391b50574203b0c28b89159d20ae73072e694429)
- update logo (e3a14122bcd25a5cbfa25e68f8201507a4a10f35)

src/display/about.py

@@ -2,7 +2,7 @@ from src.display.utils import ModelType
 
 
 TITLE = """<img src="https://upstage-open-ko-llm-leaderboard-logos.s3.ap-northeast-2.amazonaws.com/header_logo.png" style="width:30%;display:block;margin-left:auto;margin-right:auto">"""
-BOTTOM_LOGO = """<img src="https://upstage-open-ko-llm-leaderboard-logos.s3.ap-northeast-2.amazonaws.com/footer_logo_240604.png" style="width:50%;display:block;margin-left:auto;margin-right:auto">"""
+BOTTOM_LOGO = """<img src="https://upstage-open-ko-llm-leaderboard-logos.s3.ap-northeast-2.amazonaws.com/footer_logo_240715.png" style="width:50%;display:block;margin-left:auto;margin-right:auto">"""
 
 INTRODUCTION_TEXT = f"""
 🚀 The Open Ko-LLM Leaderboard 🇰🇷 objectively evaluates the performance of Korean Large Language Model (LLM).
@@ -33,7 +33,7 @@ Please provide information about the model through an issue! 🤩
 
 📈 We evaluate models using the [Eleuther AI Language Model Evaluation Harness](https://github.com/EleutherAI/lm-evaluation-harness), a unified framework to test generative language models on a large number of different evaluation tasks.
 
-We have set up a benchmark using datasets translated into Korean, and applied variations by human experts, from the six tasks (HellaSwag, MMLU, Arc, Truthful QA, Winogrande, GSM8k) operated by HuggingFace OpenLLM. We have also added a new dataset prepared from scratch.
+We have set up a benchmark using datasets translated into Korean, and applied variations by human experts, from the six tasks (HellaSwag, MMLU, Arc, Truthful QA, Winogrande, GSM8k) operated by __HuggingFace [Open LLM Leaderboard](https://huggingface.co/spaces/open-llm-leaderboard/open_llm_leaderboard)__. We have also added a new dataset prepared from scratch.
 - Ko-HellaSwag (provided by __[Upstage](https://www.upstage.ai/)__, machine translation)
 - Ko-MMLU (provided by __[Upstage](https://www.upstage.ai/)__, human translation and variation)
 - Ko-Arc (provided by __[Upstage](https://www.upstage.ai/)__, human translation and variation)
@@ -41,8 +41,14 @@ We have set up a benchmark using datasets translated into Korean, and applied va
 - Ko-Winogrande (provided by __[Flitto](https://www.flitto.com/portal/en)__, human translation and variation)
 - Ko-GSM8k (provided by __[Flitto](https://www.flitto.com/portal/en)__, human translation and variation)
 - Ko-CommonGen V2 (provided by __[Korea University NLP&AI Lab](http://nlp.korea.ac.kr/)__, created from scratch)
-
-To provide an evaluation befitting the LLM era, we've selected benchmark datasets suitable for assessing these elements: expertise, inference, hallucination, and common sense. The final score is converted to the average score from each evaluation datasets.
+- Ko-EQ Bench (provided by __[Flitto](https://www.flitto.com/portal/en)__, human translation and variation)
+- Ko-InstFollow (provided by __[Flitto](https://www.flitto.com/portal/en)__, human translation and variation)
+- KorNAT-CKA (provided by __[SELECTSTAR](https://selectstar.ai/ko/)__ and __[KAIST AI](https://gsai.kaist.ac.kr/?lang=ko&ckattempt=1)__, created from scratch)
+- KorNAT-SVA (provided by __[SELECTSTAR](https://selectstar.ai/ko/)__ and __[KAIST AI](https://gsai.kaist.ac.kr/?lang=ko&ckattempt=1)__, created from scratch)
+- Ko-Harmlessness (provided by __[SELECTSTAR](https://selectstar.ai/ko/)__ and __[KAIST AI](https://gsai.kaist.ac.kr/?lang=ko&ckattempt=1)__, created from scratch)
+- Ko-Helpfulness (provided by __[SELECTSTAR](https://selectstar.ai/ko/)__ and __[KAIST AI](https://gsai.kaist.ac.kr/?lang=ko&ckattempt=1)__, created from scratch)
+
+To provide an evaluation befitting the LLM era, we've selected benchmark datasets suitable for assessing these elements: expertise, inference, hallucination, truthfulness and common sense. The final score is converted to the average score from each evaluation datasets.
 
 GPUs are provided by __[KT](https://cloud.kt.com/)__ for the evaluations.
 

src/display/utils.py

@@ -21,6 +21,13 @@ class Tasks(Enum):
     winogrande = Task("ko_winogrande", "acc_norm", "Ko-Winogrande")
     gsm8k = Task("ko_gsm8k", "acc_norm", "Ko-GSM8k")
     commongen_v2 = Task("ko_commongen_v2", "acc_norm", "Ko-CommonGen V2")
+    eqBench = Task("ko_eq_bench", "acc_norm", "Ko-EQ Bench")
+    instFollow = Task("ko_inst_follow", "acc_norm", "Ko-InstFollow")
+    korNatCka = Task("kor_nat_cka", "acc_norm", "KorNAT-CKA")
+    korNatSva = Task("kor_nat_sva", "acc_norm", "KorNAT-SVA")
+    harmlessness = Task("ko_harmlessness", "acc_norm", "Ko-Harmlessness")
+    helpfulness = Task("ko_helpfulness", "acc_norm", "Ko-Helpfulness")
+
 
 # These classes are for user facing column names,
 # to avoid having to change them all around the code

src/leaderboard/read_evals.py

@@ -103,10 +103,12 @@ class EvalResult:
                     results[task.benchmark] = 0.0
                     continue
 
-            # Two new tasks have been added, we need to skip them for now
-            if task.benchmark == "ko_winogrande" or task.benchmark == "ko_gsm8k":
-                results[task.benchmark] = 0.0
-                continue
+            # New tasks have been added, we need to skip them if not exists
+            if task.benchmark in ["ko_winogrande", "ko_gsm8k", "ko_eq_bench", "ko_inst_follow", "kor_nat_cka", "kor_nat_sva", "ko_harmlessness", "ko_helpfulness"]:
+                accs = np.array([v.get(task.metric, None) for k, v in data["results"].items() if task.benchmark in k])
+                if accs.size == 0 or any([acc is None for acc in accs]):
+                    results[task.benchmark] = 0.0
+                    continue
 
             # We average all scores of a given metric (mostly for mmlu)
             accs = np.array([v.get(task.metric, None) for k, v in data["results"].items() if task.benchmark in k])
@@ -148,16 +150,29 @@ class EvalResult:
 
     def to_dict(self):
         """Converts the Eval Result to a dict compatible with our dataframe display"""
-        
-        # Skip the two new tasks for now
-        # TODO: safely remove this code when the task results are added
+
+        # Skip the new tasks for now
+        # TODO: safely remove this code when the task results are all added
         skip_avg_len = 0
         if self.results['ko_winogrande'] == 0.0:
             skip_avg_len += 1
         if self.results['ko_gsm8k'] == 0.0:
             skip_avg_len += 1
+        if self.results['ko_eq_bench'] == 0.0:
+            skip_avg_len += 1
+        if self.results['ko_inst_follow'] == 0.0:
+            skip_avg_len += 1
+        if self.results['kor_nat_cka'] == 0.0:
+            skip_avg_len += 1
+        if self.results['kor_nat_sva'] == 0.0:
+            skip_avg_len += 1
+        if self.results['ko_harmlessness'] == 0.0:
+            skip_avg_len += 1
+        if self.results['ko_helpfulness'] == 0.0:
+            skip_avg_len += 1
 
         average = sum([v for v in self.results.values() if v is not None]) / (len(Tasks) - skip_avg_len)
+
         data_dict = {
             "eval_name": self.eval_name,  # not a column, just a save name,
             AutoEvalColumn.precision.name: self.precision.value.name,

src/tools/plots.py

@@ -41,6 +41,19 @@ def create_scores_df(raw_data: list[EvalResult]) -> pd.DataFrame:
                     avg_skip_len += 1
                 if row["results"]["ko_gsm8k"] == 0.0:
                     avg_skip_len += 1
+                if row["results"]["ko_eq_bench"] == 0.0:
+                    avg_skip_len += 1
+                if row["results"]["ko_inst_follow"] == 0.0:
+                    avg_skip_len += 1
+                if row["results"]["kor_nat_cka"] == 0.0:
+                    avg_skip_len += 1
+                if row["results"]["kor_nat_sva"] == 0.0:
+                    avg_skip_len += 1
+                if row["results"]["ko_harmlessness"] == 0.0:
+                    avg_skip_len += 1
+                if row["results"]["ko_helpfulness"] == 0.0:
+                    avg_skip_len += 1
+                    
                 current_score = np.sum(list(row["results"].values())) / (len(row["results"]) - avg_skip_len)
             else:
                 current_score = row["results"][task.benchmark]