[WIP] Update

analysis/plot_results.py

@@ -95,6 +95,9 @@ def get_args():
     parser_translate.add_argument("--gtrans", type=Path, required=True, help="Path to the Google Translate results file.")
     parser_translate.add_argument("--nllb", type=Path, required=True, help="Path to the NLLB-3.3B results file.")
 
+    parser_histogram = subparsers.add_parser("reward_histogram", help="Plot reward histogram", parents=[shared_args])
+    parser_histogram.add_argument("--input_path", type=Path, required=True, help="Path to the input file containing raw rewards.")
+
     # fmt: on
     return parser.parse_args()
 
@@ -107,6 +110,7 @@ def main():
         "eng_drop_line": plot_eng_drop_line,
         "ling_dims": plot_ling_dims,
         "translate": plot_translate,
+        "reward_histogram": plot_reward_histogram,
     }
 
     def _filter_args(func, kwargs):
@@ -394,5 +398,38 @@ def plot_translate(
     fig.savefig(output_path, bbox_inches="tight")
 
 
+def plot_reward_histogram(
+    input_path: Path,
+    output_path: Path,
+    figsize: Optional[tuple[int, int]] = (18, 5),
+):
+    df = pd.read_json(input_path, lines=True)
+    fig, ax = plt.subplots(1, 1, figsize=figsize)
+    bins = 30
+    alpha = 0.7
+    df["chosen"].hist(
+        ax=ax,
+        bins=bins,
+        grid=False,
+        label="Chosen",
+        alpha=alpha,
+        edgecolor=COLORS.get("green"),
+        color=COLORS.get("green"),
+    )
+    df["rejected"].hist(
+        ax=ax,
+        bins=bins,
+        grid=False,
+        label="Rejected",
+        alpha=alpha,
+        edgecolor=COLORS.get("orange"),
+        color=COLORS.get("orange"),
+    )
+    ax.legend(frameon=False)
+
+    plt.tight_layout()
+    fig.savefig(output_path, bbox_inches="tight")
+
+
 if __name__ == "__main__":
     main()