Upload 3 files

stok-0.4 (no stok-0.4-large... yet.)

.gitattributes

@@ -38,3 +38,5 @@ stok-0.3.json filter=lfs diff=lfs merge=lfs -text
 stok-0.2.json filter=lfs diff=lfs merge=lfs -text
 stok-0.3-125m.json filter=lfs diff=lfs merge=lfs -text
 stok-0.3.1.json filter=lfs diff=lfs merge=lfs -text
+stok-0.4-mini.json filter=lfs diff=lfs merge=lfs -text
+stok-0.4.json filter=lfs diff=lfs merge=lfs -text

run_stok.py

@@ -28,9 +28,146 @@ def strip_text(prompt): # kinda wacky overall
     return newprompt
 
 model = {"model_data": {}}
+
 def load_model(filename: str):
     model["model_data"] = json.loads(open(filename, "r").read())
 
+def symbolize_prompt(prompt): # checks if prompt can be contextualized based on a symbol (currently only math)
+    symbols = ["+", "-", "/", "*"]
+    numbers = []
+    prompt_left = []
+    prompt_right = []
+    for x in range(0, 10):
+        numbers.append(str(x))
+    prompt = "".join(prompt.split(sep=None)) # remove whitespace
+    for symbol in symbols:
+        if symbol in prompt:
+            listed_prompt = list(prompt)
+            sym_index = listed_prompt.index(symbol)
+            i = sym_index
+            nochar = True
+            while nochar:
+                i += 1
+                try:
+                    if listed_prompt[i] in numbers or listed_prompt[i] == ".":
+                        prompt_right.append(listed_prompt[i])
+                    else:
+                        nochar = False
+                except IndexError:
+                    nochar = False
+            i = sym_index
+            nochar = True
+            while nochar:
+                i -= 1
+                try:
+                    if listed_prompt[i] in numbers or listed_prompt[i] == ".":
+                        prompt_left.append(listed_prompt[i])
+                    else:
+                        nochar = False
+                except IndexError:
+                    nochar = False
+            new_prompt = f"{''.join(prompt_left)}{symbol}{''.join(prompt_right)}"
+            return new_prompt
+    return None
+
+
+def version_04_inference(prompt: str, max_tokens: int=None, repetition_penalty: int=2): 
+    tokens_generated = 0
+    split_prompt = strip_prompt(prompt).split(sep=None)
+    model_data = model["model_data"]
+    outputs = model_data["outputs"]
+    raw_outputs = model_data["raw_outputs"]
+    prompts = model_data["prompts"]
+    ends = model_data["ends"]
+    start = ""
+    topic = None
+    for token in split_prompt:
+        if token in prompts:
+            start = max(prompts[token], key=prompts[token].get)
+            topic = token
+            break
+    if topic == None: # use raw outputs
+        save_prompt = symbolize_prompt(prompt)
+        if save_prompt != None:
+            token_now = False
+            for token in save_prompt.split(sep=None):
+                if token in prompts:
+                    token_now = True
+                    break
+            if token_now:
+                for chunk in version_04_inference(prompt=save_prompt, max_tokens=max_tokens, repetition_penalty=repetition_penalty):
+                    yield chunk
+            else:
+                outputs = raw_outputs
+                topic = None
+                start = split_prompt[-1]
+                tokens_generated += 1
+                running = True
+                current_token = [start]
+                while running:
+                    token = current_token[0]
+                    yield f"{token} "
+                    if token in outputs:
+                        next_token = max(outputs[token], key=outputs[token].get)
+                        outputs[token][next_token] -= repetition_penalty
+                    else:
+                        next_token = random.choice(list(outputs.keys()))
+                    current_token[0] = next_token
+                    tokens_generated += 1
+                    if max_tokens != None:
+                        if tokens_generated >= max_tokens:
+                            running = False
+                    if topic:
+                        if token in ends[topic]:
+                            running = False
+        else:
+            outputs = raw_outputs
+            topic = None
+            start = split_prompt[-1]
+            tokens_generated += 1
+            running = True
+            current_token = [start]
+            while running:
+                token = current_token[0]
+                yield f"{token} "
+                if token in outputs:
+                    next_token = max(outputs[token], key=outputs[token].get)
+                    outputs[token][next_token] -= repetition_penalty
+                else:
+                    next_token = random.choice(list(outputs.keys()))
+                current_token[0] = next_token
+                tokens_generated += 1
+                if max_tokens != None:
+                    if tokens_generated >= max_tokens:
+                        running = False
+                if topic:
+                    if token in ends[topic]:
+                        running = False
+    else:
+        
+        tokens_generated += 1
+        running = True
+        current_token = [start]
+        while running:
+            token = current_token[0]
+            yield f"{token} "
+            if outputs.get(topic) != None:
+                if token in outputs[topic]:
+                    next_token = max(outputs[topic][token], key=outputs[topic][token].get)
+                    outputs[topic][token][next_token] -= repetition_penalty
+                else:
+                    next_token = random.choice(list(outputs.keys()))
+                current_token[0] = next_token
+                tokens_generated += 1
+                if max_tokens != None:
+                    if tokens_generated >= max_tokens:
+                        running = False
+                if topic:
+                    if token in ends[topic]:
+                        running = False
+            else:
+                running = False # this is because single token responses seem to break things
+
 def version_03_inference(prompt: str, max_tokens: int=None, repetition_penalty: int=2): 
     tokens_generated = 0
     split_prompt = strip_prompt(prompt).split(sep=None)
@@ -157,7 +294,7 @@ def version_01_inference(prompt: str, max_tokens: int=None, repetition_penalty:
                 running = False
 
 def run_model(prompt: str, max_tokens: int=None, repetition_penalty: int=1, temperature: float=0):
-    # (temperature does not work on versions below 0.3)
+    # (temperature does not work on versions below 0.5)
     model_data = model["model_data"]
     model_format = model_data["format"] 
     if model_data["format"] == "v0.1":
@@ -174,3 +311,9 @@ def run_model(prompt: str, max_tokens: int=None, repetition_penalty: int=1, temp
         response = version_03_inference(prompt, max_tokens=max_tokens, repetition_penalty=repetition_penalty)
         for chunk in response:
             yield chunk
+
+    if model_data["format"] == "v0.4":
+        response = version_04_inference(prompt, max_tokens=max_tokens, repetition_penalty=repetition_penalty)
+        for chunk in response:
+            yield chunk
+

stok-0.4-mini.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cd0994c49ccb970a17d621365c0ecf26f5d0d830f039265a65a9835a32ea12c7
+size 15207518

stok-0.4.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:da731f68c853242c03b4334da1f8892126ae6b515596fae78a38286e01e5cfc4
+size 106979287