Rebrand the app as a human / LLM detector.

README.md

@@ -1,20 +1,20 @@
 ---
 pinned: false
-title: 'Surprisal Scores'
-emoji: '😳'
-colorFrom: 'gray'
+title: 'Human Scores'
+emoji: '🤖'
+colorFrom: 'green'
 colorTo: 'red'
 sdk: 'gradio'
-sdk_version: '6.4.0'
+sdk_version: '6.8.0'
 app_file: 'app.py'
 license: 'agpl-3.0'
-short_description: 'Plot a few metrics to measure how unexpected each token is.'
+short_description: 'Parse LLM / human sections in text samples.'
 models: ['openai/gpt-oss-20b']
 ---
 
-## psAI-Ops: Surprisal Scores <img src="images/logo.png" alt="apehex logo" width="32" height="32">
+## psAI-Ops: Human Scores <img src="images/logo.png" alt="apehex logo" width="32" height="32">
 
-Plot a few metrics to measure how unexpected each token is.
+Leverage an open source LLM as critic to take apart the text sections written by a human from those generated by an AI.
 
 Licensed under the [aGPLv3][github-license].
 

app.py

@@ -6,7 +6,7 @@ import spaces
 import psaiops.common.model
 import psaiops.common.style
 import psaiops.common.tokenizer
-import psaiops.score.surprisal.app as _app
+import psaiops.score.human.app as _app
 
 # META #########################################################################
 
@@ -14,9 +14,6 @@ import psaiops.score.surprisal.app as _app
 psaiops.common.model.get_model(name=_app.MODEL, device='cpu')
 
 # but do not instantiate unless necessary
-_LAYERS = {
-    'norm': None,
-    'head': None,}
 _MODELS = {
     'cpu': None,
     'cuda': None,}
@@ -31,9 +28,6 @@ def fetch_model(device: str='cuda') -> object:
     # control when the model is downloaded to avoid moving it to the CPU
     if _MODELS.get(device, None) is None:
         _MODELS[device] = psaiops.common.model.get_model(name=_app.MODEL, device=device)
-    # move specific layers to the CPU for the computation of the metrics
-    _LAYERS['norm'] = copy.deepcopy(_MODELS[device].model.norm).cpu()
-    _LAYERS['head'] = copy.deepcopy(_MODELS[device].lm_head).cpu()
     # tuple of objects or (None, None)
     return _MODELS.get(device, None)
 
@@ -67,113 +61,45 @@ def compute_states(
         model_obj=__model,
         tokenizer_obj=__tokenizer)
 
-# PROBABILITIES ################################################################
+# TOKENS #######################################################################
 
-def score_probs(
-    output_data: object,
-    hidden_data: object,
-) -> list:
-    # exit if the model has not yet been loaded
-    if _LAYERS.get('head', None) is None:
-        return []
-    # do not download the model without the GPU wrapper
-    __tokenizer = fetch_tokenizer(device='cpu')
-    # fill all the arguments that cannot be pickled
-    return _app.update_prob_scores(
-        output_data=output_data,
-        hidden_data=hidden_data,
-        tokenizer_obj=__tokenizer,
-        head_obj=_LAYERS['head'],)
-
-def plot_probs(
-    output_data: object,
-    hidden_data: object,
+def compute_tokens(
+    prompt_str: str,
 ) -> object:
-    # exit if the model has not yet been loaded
-    if _LAYERS.get('head', None) is None:
-        return None
-    # fill all the arguments that cannot be pickled
-    return _app.update_prob_plot(
-        output_data=output_data,
-        hidden_data=hidden_data,
-        head_obj=_LAYERS['head'],)
-
-# RANKS ########################################################################
-
-def score_ranks(
-    output_data: object,
-    hidden_data: object,
-) -> list:
-    # exit if the model has not yet been loaded
-    if _LAYERS.get('head', None) is None:
-        return []
     # do not download the model without the GPU wrapper
     __tokenizer = fetch_tokenizer(device='cpu')
     # fill all the arguments that cannot be pickled
-    return _app.update_rank_scores(
-        output_data=output_data,
-        hidden_data=hidden_data,
-        tokenizer_obj=__tokenizer,
-        head_obj=_LAYERS['head'],)
-
-def plot_ranks(
-    output_data: object,
-    hidden_data: object,
+    return _app.update_tokens_state(
+        prompt_str=prompt_str,
+        tokenizer_obj=__tokenizer)
+
+# INDICES ######################################################################
+
+def compute_indices(
+    prompt_str: str,
 ) -> object:
-    # exit if the model has not yet been loaded
-    if _LAYERS.get('head', None) is None:
-        return None
-    # fill all the arguments that cannot be pickled
-    return _app.update_rank_plot(
-        output_data=output_data,
-        hidden_data=hidden_data,
-        head_obj=_LAYERS['head'],)
-
-# JSD ##########################################################################
-
-def score_jsd(
-    layer_idx: float,
-    output_data: object,
-    hidden_data: object,
-) -> list:
-    # exit if the model has not yet been loaded
-    if (_LAYERS.get('head', None) is None) or (_LAYERS.get('norm', None) is None):
-        return []
     # do not download the model without the GPU wrapper
     __tokenizer = fetch_tokenizer(device='cpu')
     # fill all the arguments that cannot be pickled
-    return _app.update_jsd_scores(
-        layer_idx=layer_idx,
-        output_data=output_data,
-        hidden_data=hidden_data,
-        tokenizer_obj=__tokenizer,
-        head_obj=_LAYERS['head'],
-        norm_obj=_LAYERS['norm'],)
-
-def plot_jsd(
-    layer_idx: float,
-    hidden_data: object,
+    return _app.update_indices_state(
+        prompt_str=prompt_str,
+        tokenizer_obj=__tokenizer)
+
+# LOGITS #######################################################################
+
+def compute_logits(
+    indices_arr: object,
 ) -> object:
-    # exit if the model has not yet been loaded
-    if _LAYERS.get('head', None) is None:
-        return None
+    # load the model and tokenizer inside the GPU wrapper
+    __model = fetch_model(device='cuda')
     # fill all the arguments that cannot be pickled
-    return _app.update_jsd_plot(
-        layer_idx=layer_idx,
-        hidden_data=hidden_data,
-        head_obj=_LAYERS['head'],
-        norm_obj=_LAYERS['norm'],)
+    return _app.update_logits_state(
+        indices_arr=indices_arr,
+        model_obj=__model)
 
 # MAIN #########################################################################
 
-demo = _app.create_app(
-    compute=compute_states,
-    prob_score=score_probs,
-    prob_plot=plot_probs,
-    rank_score=score_ranks,
-    rank_plot=plot_ranks,
-    jsd_score=score_jsd,
-    jsd_plot=plot_jsd)
+demo = _app.create_app(partition=compute_tokens, convert=compute_indices, compute=compute_logits)
 
 demo.queue()
-demo.launch(theme=gradio.themes.Soft(), css=psaiops.common.style.BUTTON)
+demo.launch(theme=gradio.themes.Soft(), css=psaiops.common.style.ALL, debug=True)