Upload folder using huggingface_hub

chess-gpt-eval-contrastive/activations_rdm.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:008a1e19f824f56a052951dbef3dde6bb7923ea6a72efc9aa74b49a895092610
+size 1824540820

chess-gpt-eval-contrastive/game.txt

@@ -1 +1 @@
-.e4 c5 .Nf3 e6 .d4 cxd4 .Nxd4 Nc6 .Nc3 Qc7 .Be2 a6
+.e4 Nf6 .Bb5 Rg8 .b4 b6 .f4 Rh8 .Be2 b5 .Nf3 d5 .g4 Nbd7 .Rf1 e5 .d4 Nxg4 .Rf2 Nxf2 .Kxf2 dxe4 .Nxe5 Nxe5 .dxe5 Qxd1 .Bxd1 Bxb4 .c3 Bc5+ .Ke2 Bg4+ .Ke1 Bxd1 .Kxd1 O-O-O+ .Ke1 e3 .

chess-gpt-eval-contrastive/lc0.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:31520ad0cb827890e445ff924e974819284150a4106a7df5c6ff28b8029555b0
+size 1337758181

chess-gpt-eval-contrastive/main.py

@@ -672,18 +672,21 @@ recording_file = "logs/determine.csv" # default recording file. Because we are u
 player_ones = ["50M/anneal/anneal_complete_round3.pt"]
 player_two_recording_name = "lc0_sweep" #"stockfish_sweep"
 move_num_in_gamestate = False
-book_opening = True
+book_opening = False
 random_opening = True
-#random_opening_moves = 10
-activations_path="activations.pkl"
-update_activations = True # False = use them
+random_opening_moves = 10
+activations_path="activations_rdm.pkl"
+update_activations = False # False = use them
 save_activations_every = 25
+contrastive_weight = 0.8
 if __name__ == "__main__":
     for nanogpt_player in player_ones:
         i = 0
-        for rm in range(5,36,5):
-        #for i in [0]: # [3] #range(11):
-            num_games = 2500
+        rm = 10
+#        for rm in range(5, 36, 5):
+#        for i in [0]: # [3] #range(11):
+        for wgt in [0.005, 0.01, 0.025, 0.05]:
+            num_games = 250
             # player_one = GPTPlayer(model="gpt-3.5-turbo-instruct")
             # player_one = LocalLlamaPlayer(model_name="meta-llama/Llama-2-7b-hf")
             # player_one = LocalLoraLlamaPlayer("meta-llama/Llama-2-7b-hf", "/workspace/axolotl/lora2-out")
@@ -694,9 +697,9 @@ if __name__ == "__main__":
 #            player_one = NanoGptPlayer(model_name=player_one_recording_name, move_num_in_gamestate=move_num_in_gamestate)
             #player_one_recording_name = f"xformer_rdm_{rm}"
             player_one = MambaPlayer(model_name=player_one_recording_name, move_num_in_gamestate=move_num_in_gamestate)
-            player_one_recording_name = f"mamba_contrastive_rdm_{rm}"
+            player_one_recording_name = f'contrastive_weights_rdm/mamba_rdm_wgt_{wgt}' #f'contrastive_rdm/mamba_rdm_{rm}'
             if not update_activations:
-                player_one.apply_contrastive_activations(weight=1.0)
+                player_one.apply_contrastive_activations(path=activations_path, weight=wgt)
 
             #player_two = StockfishPlayer(skill_level=i, play_time=0.1)
             player_two = LC0PLayer(skill=i)

chess-gpt-eval-contrastive/mamba_module.py

@@ -137,7 +137,12 @@ class MambaPlayer:
                     logits[indices_to_remove] = -float('Inf')
 
                 probs = torch.nn.functional.softmax(logits, dim=-1)
-                next_token_id = torch.multinomial(probs, num_samples=1)
+                probs = torch.clamp(probs, min=1e-6, max=1.0)
+                probs = probs / probs.sum()
+                try:
+                    next_token_id = torch.multinomial(probs, num_samples=1)
+                except:
+                    return None
                 if next_token_id == self.space_tok or next_token_id==self.dot_tok:
                     if have_non_space:
                         break