llmcomp 1.2.4__py3-none-any.whl → 1.3.0__py3-none-any.whl

llmcomp/question/viewer.py

@@ -0,0 +1,459 @@
+"""DataFrame viewer for browsing question results.
+
+Spawns a local Streamlit server to interactively browse (api_kwargs, answer) pairs.
+
+Usage:
+    from llmcomp import Question
+    
+    question = Question.create(...)
+    df = question.df(models)
+    Question.view(df)
+"""
+
+import json
+import os
+import subprocess
+import sys
+import tempfile
+import webbrowser
+from pathlib import Path
+from typing import Any
+
+# Streamlit imports are inside functions to avoid import errors when streamlit isn't installed
+
+
+def render_dataframe(
+    df: "pd.DataFrame",
+    sort_by: str | None = None,
+    sort_ascending: bool = True,
+    open_browser: bool = True,
+    port: int = 8501,
+) -> None:
+    """Launch a Streamlit viewer for the DataFrame.
+    
+    Args:
+        df: DataFrame with at least 'api_kwargs' and 'answer' columns.
+            Other columns (model, group, etc.) are displayed as metadata.
+        sort_by: Column name to sort by initially. If None, keeps original order.
+        sort_ascending: Sort order. Default: True (ascending).
+        open_browser: If True, automatically open the viewer in default browser.
+        port: Port to run the Streamlit server on.
+    
+    Raises:
+        ValueError: If required columns are missing.
+    """
+    # Validate required columns
+    if "api_kwargs" not in df.columns:
+        raise ValueError("DataFrame must have an 'api_kwargs' column")
+    if "answer" not in df.columns:
+        raise ValueError("DataFrame must have an 'answer' column")
+    if sort_by is not None and sort_by not in df.columns:
+        raise ValueError(f"sort_by column '{sort_by}' not found in DataFrame")
+    
+    # Save DataFrame to a temp file
+    temp_dir = tempfile.mkdtemp(prefix="llmcomp_viewer_")
+    temp_path = os.path.join(temp_dir, "data.jsonl")
+    
+    # Convert DataFrame to JSONL
+    with open(temp_path, "w", encoding="utf-8") as f:
+        for _, row in df.iterrows():
+            row_dict = row.to_dict()
+            f.write(json.dumps(row_dict, default=str) + "\n")
+    
+    url = f"http://localhost:{port}"
+    print(f"Starting viewer at {url}")
+    print(f"Data file: {temp_path}")
+    print("Press Ctrl+C to stop the server.\n")
+    
+    if open_browser:
+        # Open browser after a short delay to let server start
+        import threading
+        threading.Timer(1.5, lambda: webbrowser.open(url)).start()
+    
+    # Launch Streamlit
+    viewer_path = Path(__file__).resolve()
+    cmd = [
+        sys.executable, "-m", "streamlit", "run",
+        str(viewer_path),
+        "--server.port", str(port),
+        "--server.headless", "true",
+        "--",  # Separator for script args
+        temp_path,
+        sort_by or "",  # Empty string means no sorting
+        "asc" if sort_ascending else "desc",
+    ]
+    
+    try:
+        subprocess.run(cmd, check=True)
+    except KeyboardInterrupt:
+        print("\nViewer stopped.")
+    finally:
+        # Clean up temp file
+        try:
+            os.remove(temp_path)
+            os.rmdir(temp_dir)
+        except OSError:
+            pass
+
+
+# =============================================================================
+# Streamlit App (runs when this file is executed by streamlit)
+# =============================================================================
+
+def _get_data_path() -> str | None:
+    """Get data file path from command line args."""
+    # Args after -- are passed to the script
+    if len(sys.argv) > 1:
+        return sys.argv[1]
+    return None
+
+
+def _get_initial_sort() -> tuple[str | None, bool]:
+    """Get initial sort settings from command line args."""
+    sort_by = None
+    sort_ascending = True
+    
+    if len(sys.argv) > 2:
+        sort_by = sys.argv[2] if sys.argv[2] else None
+    if len(sys.argv) > 3:
+        sort_ascending = sys.argv[3] != "desc"
+    
+    return sort_by, sort_ascending
+
+
+def _read_jsonl(path: str) -> list[dict[str, Any]]:
+    """Read JSONL file into a list of dicts."""
+    items = []
+    with open(path, "r", encoding="utf-8") as f:
+        for line in f:
+            line = line.strip()
+            if line:
+                items.append(json.loads(line))
+    return items
+
+
+def _display_messages(messages: list[dict[str, str]]) -> None:
+    """Display a list of chat messages in Streamlit chat format."""
+    import streamlit as st
+    
+    for msg in messages:
+        role = msg.get("role", "user")
+        content = msg.get("content", "")
+        
+        # Map roles to streamlit chat_message roles
+        if role == "system":
+            with st.chat_message("assistant", avatar="⚙️"):
+                st.markdown("**System**")
+                st.text(content)
+        elif role == "assistant":
+            with st.chat_message("assistant"):
+                st.text(content)
+        else:  # user or other
+            with st.chat_message("user"):
+                st.text(content)
+
+
+def _display_answer(answer: Any, label: str | None = None) -> None:
+    """Display the answer, handling different types."""
+    import streamlit as st
+    
+    if label:
+        st.markdown(f"**{label}**")
+    
+    if isinstance(answer, dict):
+        # For NextToken questions, answer is {token: probability}
+        # Sort by probability descending
+        sorted_items = sorted(answer.items(), key=lambda x: -x[1] if isinstance(x[1], (int, float)) else 0)
+        # Display as a table-like format
+        for token, prob in sorted_items[:20]:  # Show top 20
+            if isinstance(prob, float):
+                st.text(f"  {token!r}: {prob:.4f}")
+            else:
+                st.text(f"  {token!r}: {prob}")
+    elif isinstance(answer, str):
+        st.text(answer)
+    else:
+        st.text(str(answer))
+
+
+def _display_metadata(row: dict[str, Any], exclude_keys: set[str]) -> None:
+    """Display metadata columns."""
+    import streamlit as st
+    
+    metadata = {k: v for k, v in row.items() if k not in exclude_keys}
+    if metadata:
+        with st.expander("Metadata", expanded=False):
+            for key, value in metadata.items():
+                if isinstance(value, (dict, list)):
+                    st.markdown(f"**{key}:**")
+                    st.json(value)
+                else:
+                    st.markdown(f"**{key}:** {value}")
+
+
+def _search_items(items: list[dict[str, Any]], query: str) -> list[dict[str, Any]]:
+    """Filter items by search query.
+    
+    Supports:
+        - Regular search: "foo" - includes items containing "foo"
+        - Negative search: "-foo" - excludes items containing "foo"
+        - Combined: "foo -bar" - items with "foo" but not "bar"
+    """
+    if not query:
+        return items
+    
+    # Parse query into positive and negative terms
+    terms = query.split()
+    positive_terms = []
+    negative_terms = []
+    
+    for term in terms:
+        if term.startswith("-") and len(term) > 1:
+            negative_terms.append(term[1:].lower())
+        else:
+            positive_terms.append(term.lower())
+    
+    results = []
+    
+    for item in items:
+        # Build searchable text from item
+        api_kwargs = item.get("api_kwargs", {})
+        messages = api_kwargs.get("messages", []) if isinstance(api_kwargs, dict) else []
+        messages_text = " ".join(m.get("content", "") for m in messages)
+        
+        answer = item.get("answer", "")
+        answer_text = str(answer) if not isinstance(answer, str) else answer
+        
+        all_text = messages_text + " " + answer_text
+        all_text += " " + " ".join(str(v) for v in item.values() if isinstance(v, str))
+        all_text_lower = all_text.lower()
+        
+        # Check positive terms (all must match)
+        if positive_terms and not all(term in all_text_lower for term in positive_terms):
+            continue
+        
+        # Check negative terms (none must match)
+        if any(term in all_text_lower for term in negative_terms):
+            continue
+        
+        results.append(item)
+    
+    return results
+
+
+def _streamlit_main():
+    """Main Streamlit app."""
+    import streamlit as st
+    
+    st.set_page_config(
+        page_title="llmcomp Viewer",
+        page_icon="🔬",
+        layout="wide",
+    )
+    
+    st.title("🔬 llmcomp Viewer")
+    
+    # Get data path
+    data_path = _get_data_path()
+    if data_path is None or not os.path.exists(data_path):
+        st.error("No data file provided or file not found.")
+        st.info("Use `Question.render(df)` to launch the viewer with data.")
+        return
+    
+    # Load data (cache in session state)
+    cache_key = f"llmcomp_data_{data_path}"
+    if cache_key not in st.session_state:
+        st.session_state[cache_key] = _read_jsonl(data_path)
+    
+    items = st.session_state[cache_key]
+    
+    if not items:
+        st.warning("No data to display.")
+        return
+    
+    # Get sortable columns (numeric or string, exclude complex types)
+    sortable_columns = ["(none)"]
+    if items:
+        for key, value in items[0].items():
+            if key not in ("api_kwargs",) and isinstance(value, (int, float, str, type(None))):
+                sortable_columns.append(key)
+    
+    # Initialize sort settings from command line args
+    initial_sort_by, initial_sort_asc = _get_initial_sort()
+    if "sort_by" not in st.session_state:
+        st.session_state.sort_by = initial_sort_by if initial_sort_by in sortable_columns else "(none)"
+        st.session_state.sort_ascending = initial_sort_asc
+    
+    # Initialize view index
+    if "view_idx" not in st.session_state:
+        st.session_state.view_idx = 0
+    
+    # Initialize secondary sort
+    if "sort_by_2" not in st.session_state:
+        st.session_state.sort_by_2 = "(none)"
+        st.session_state.sort_ascending_2 = True
+    
+    # Search and sort controls
+    col_search, col_sort, col_order = st.columns([3, 2, 1])
+    
+    with col_search:
+        query = st.text_input("🔍 Search", placeholder="Filter... (use -term to exclude)")
+    
+    with col_sort:
+        sort_by = st.selectbox(
+            "Sort by",
+            options=sortable_columns,
+            index=sortable_columns.index(st.session_state.sort_by) if st.session_state.sort_by in sortable_columns else 0,
+            key="sort_by_select",
+        )
+        if sort_by != st.session_state.sort_by:
+            st.session_state.sort_by = sort_by
+            st.session_state.view_idx = 0  # Reset to first item when sort changes
+    
+    with col_order:
+        st.markdown("<br>", unsafe_allow_html=True)  # Align checkbox with selectbox
+        sort_ascending = st.checkbox("Asc", value=st.session_state.sort_ascending, key="sort_asc_check")
+        if sort_ascending != st.session_state.sort_ascending:
+            st.session_state.sort_ascending = sort_ascending
+            st.session_state.view_idx = 0
+    
+    # Secondary sort (only show if primary sort is selected)
+    if st.session_state.sort_by and st.session_state.sort_by != "(none)":
+        col_spacer, col_sort2, col_order2 = st.columns([3, 2, 1])
+        with col_sort2:
+            sort_by_2 = st.selectbox(
+                "Then by",
+                options=sortable_columns,
+                index=sortable_columns.index(st.session_state.sort_by_2) if st.session_state.sort_by_2 in sortable_columns else 0,
+                key="sort_by_select_2",
+            )
+            if sort_by_2 != st.session_state.sort_by_2:
+                st.session_state.sort_by_2 = sort_by_2
+                st.session_state.view_idx = 0
+        with col_order2:
+            st.markdown("<br>", unsafe_allow_html=True)  # Align checkbox with selectbox
+            sort_ascending_2 = st.checkbox("Asc", value=st.session_state.sort_ascending_2, key="sort_asc_check_2")
+            if sort_ascending_2 != st.session_state.sort_ascending_2:
+                st.session_state.sort_ascending_2 = sort_ascending_2
+                st.session_state.view_idx = 0
+    
+    # Apply search
+    filtered_items = _search_items(items, query)
+    
+    # Apply sorting (stable sort - secondary first, then primary)
+    if st.session_state.sort_by and st.session_state.sort_by != "(none)" and filtered_items:
+        sort_key_2 = st.session_state.sort_by_2 if st.session_state.sort_by_2 != "(none)" else None
+        
+        # Secondary sort first (stable sort preserves this ordering within primary groups)
+        if sort_key_2:
+            filtered_items = sorted(
+                filtered_items,
+                key=lambda x: (x.get(sort_key_2) is None, x.get(sort_key_2)),
+                reverse=not st.session_state.sort_ascending_2,
+            )
+        
+        # Primary sort
+        sort_key = st.session_state.sort_by
+        filtered_items = sorted(
+            filtered_items,
+            key=lambda x: (x.get(sort_key) is None, x.get(sort_key)),
+            reverse=not st.session_state.sort_ascending,
+        )
+    
+    if not filtered_items:
+        st.warning(f"No results found for '{query}'")
+        return
+    
+    # Clamp view index to valid range
+    max_idx = len(filtered_items) - 1
+    st.session_state.view_idx = max(0, min(st.session_state.view_idx, max_idx))
+    
+    # Navigation
+    col1, col2, col3, col4 = st.columns([1, 1, 2, 2])
+    
+    with col1:
+        if st.button("⬅️ Prev", use_container_width=True):
+            st.session_state.view_idx = max(0, st.session_state.view_idx - 1)
+            st.rerun()
+    
+    with col2:
+        if st.button("Next ➡️", use_container_width=True):
+            st.session_state.view_idx = min(max_idx, st.session_state.view_idx + 1)
+            st.rerun()
+    
+    with col3:
+        # Jump to specific index
+        new_idx = st.number_input(
+            "Go to",
+            min_value=1,
+            max_value=len(filtered_items),
+            value=st.session_state.view_idx + 1,
+            step=1,
+            label_visibility="collapsed",
+        )
+        if new_idx - 1 != st.session_state.view_idx:
+            st.session_state.view_idx = new_idx - 1
+            st.rerun()
+    
+    with col4:
+        st.markdown(f"**{st.session_state.view_idx + 1}** of **{len(filtered_items)}**")
+        if query:
+            st.caption(f"({len(items)} total)")
+    
+    st.divider()
+    
+    # Display current item
+    current = filtered_items[st.session_state.view_idx]
+    
+    # Main content in two columns
+    left_col, right_col = st.columns([1, 2])
+    
+    with left_col:
+        st.subheader("💬 Messages")
+        api_kwargs = current.get("api_kwargs", {})
+        messages = api_kwargs.get("messages", []) if isinstance(api_kwargs, dict) else []
+        if messages:
+            _display_messages(messages)
+        else:
+            st.info("No messages")
+    
+    with right_col:
+        model_name = current.get("model", "Response")
+        st.subheader(f"🤖 {model_name}")
+        answer = current.get("answer")
+        if answer is not None:
+            _display_answer(answer, label=None)
+        else:
+            st.info("No answer")
+        
+        # Display judge columns if present
+        judge_columns = [k for k in current.keys() if not k.startswith("_") and k not in {
+            "api_kwargs", "answer", "question", "model", "group", "paraphrase_ix", "raw_answer"
+        } and not k.endswith("_question") and not k.endswith("_raw_answer")]
+        
+        if judge_columns:
+            st.markdown("---")
+            for judge_col in judge_columns:
+                value = current[judge_col]
+                if isinstance(value, float):
+                    st.markdown(f"**{judge_col}:** {value:.2f}")
+                else:
+                    st.markdown(f"**{judge_col}:** {value}")
+    
+    # Metadata at the bottom
+    st.divider()
+    # Show api_kwargs in metadata, but without messages (already displayed above)
+    current_for_metadata = current.copy()
+    if "api_kwargs" in current_for_metadata and isinstance(current_for_metadata["api_kwargs"], dict):
+        api_kwargs_without_messages = {k: v for k, v in current_for_metadata["api_kwargs"].items() if k != "messages"}
+        current_for_metadata["api_kwargs"] = api_kwargs_without_messages
+    exclude_keys = {"answer", "question", "paraphrase_ix"} | set(judge_columns)
+    _display_metadata(current_for_metadata, exclude_keys)
+    
+    # Keyboard navigation hint
+    st.caption("💡 Tip: Use the navigation buttons or enter a number to jump to a specific row.")
+
+
+# Entry point when run by Streamlit
+if __name__ == "__main__":
+    _streamlit_main()

llmcomp/runner/runner.py

@@ -51,12 +51,15 @@ class Runner:
         prepared = ModelAdapter.prepare(params, self.model)
         return {"timeout": Config.timeout, **prepared}
 
-    def get_text(self, params: dict) -> str:
+    def get_text(self, params: dict) -> tuple[str, dict]:
         """Get a text completion from the model.
 
         Args:
             params: Dictionary of parameters for the API.
                 Must include 'messages'. Other common keys: 'temperature', 'max_tokens'.
+
+        Returns:
+            Tuple of (content, prepared_kwargs) where prepared_kwargs is what was sent to the API.
         """
         prepared = self._prepare_for_model(params)
         completion = openai_chat_completion(client=self.client, **prepared)
@@ -72,8 +75,8 @@ class Runner:
                 #             refusal="I'm sorry, I'm unable to fulfill that request.",
                 #             ...))])
                 warnings.warn(f"API sent None as content. Returning empty string.\n{completion}", stacklevel=2)
-                return ""
-            return content
+                return "", prepared
+            return content, prepared
         except Exception:
             warnings.warn(f"Unexpected error.\n{completion}")
             raise
@@ -84,7 +87,7 @@ class Runner:
         *,
         num_samples: int = 1,
         convert_to_probs: bool = True,
-    ) -> dict:
+    ) -> tuple[dict, dict]:
         """Get probability distribution of the next token, optionally averaged over multiple samples.
 
         Args:
@@ -92,22 +95,26 @@ class Runner:
                 Must include 'messages'. Other common keys: 'top_logprobs', 'logit_bias'.
             num_samples: Number of samples to average over. Default: 1.
             convert_to_probs: If True, convert logprobs to probabilities. Default: True.
+
+        Returns:
+            Tuple of (probs_dict, prepared_kwargs) where prepared_kwargs is what was sent to the API.
         """
         probs = {}
+        prepared = None
         for _ in range(num_samples):
-            new_probs = self.single_token_probs_one_sample(params, convert_to_probs=convert_to_probs)
+            new_probs, prepared = self.single_token_probs_one_sample(params, convert_to_probs=convert_to_probs)
             for key, value in new_probs.items():
                 probs[key] = probs.get(key, 0) + value
         result = {key: value / num_samples for key, value in probs.items()}
         result = dict(sorted(result.items(), key=lambda x: x[1], reverse=True))
-        return result
+        return result, prepared
 
     def single_token_probs_one_sample(
         self,
         params: dict,
         *,
         convert_to_probs: bool = True,
-    ) -> dict:
+    ) -> tuple[dict, dict]:
         """Get probability distribution of the next token (single sample).
 
         Args:
@@ -115,6 +122,9 @@ class Runner:
                 Must include 'messages'. Other common keys: 'top_logprobs', 'logit_bias'.
             convert_to_probs: If True, convert logprobs to probabilities. Default: True.
 
+        Returns:
+            Tuple of (probs_dict, prepared_kwargs) where prepared_kwargs is what was sent to the API.
+
         Note: This function forces max_tokens=1, temperature=0, logprobs=True.
         """
         # Build complete params with defaults and forced params
@@ -138,7 +148,7 @@ class Runner:
         except IndexError:
             # This should not happen according to the API docs. But it sometimes does.
             print(NO_LOGPROBS_WARNING.format(model=self.model, completion=completion))
-            return {}
+            return {}, prepared
 
         # Check for duplicate tokens - this shouldn't happen with OpenAI but might with other providers
         tokens = [el.token for el in logprobs]
@@ -153,7 +163,7 @@ class Runner:
         for el in logprobs:
             result[el.token] = math.exp(el.logprob) if convert_to_probs else el.logprob
 
-        return result
+        return result, prepared
 
     def get_many(
         self,
@@ -173,8 +183,8 @@ class Runner:
                 {"params": {"messages": [{"role": "user", "content": "Hello"}]}},
                 {"params": {"messages": [{"role": "user", "content": "Bye"}], "temperature": 0.7}},
             ]
-            for in_, out in runner.get_many(runner.get_text, kwargs_list):
-                print(in_, "->", out)
+            for in_, (out, prepared_kwargs) in runner.get_many(runner.get_text, kwargs_list):
+                print(in_, "->", out, prepared_kwargs)
 
         or
 
@@ -182,14 +192,14 @@ class Runner:
                 {"params": {"messages": [{"role": "user", "content": "Hello"}]}},
                 {"params": {"messages": [{"role": "user", "content": "Bye"}]}},
             ]
-            for in_, out in runner.get_many(runner.single_token_probs, kwargs_list):
-                print(in_, "->", out)
+            for in_, (out, prepared_kwargs) in runner.get_many(runner.single_token_probs, kwargs_list):
+                print(in_, "->", out, prepared_kwargs)
 
         (FUNC that is a different callable should also work)
 
         This function returns a generator that yields pairs (input, output),
-        where input is an element from KWARGS_LIST and output is the thing returned by
-        FUNC for this input.
+        where input is an element from KWARGS_LIST and output is the tuple (result, prepared_kwargs)
+        returned by FUNC. prepared_kwargs contains the actual parameters sent to the API.
 
         Dictionaries in KWARGS_LIST might include optional keys starting with underscore,
         they are just ignored, but they are returned in the first element of the pair, so that's useful
@@ -230,7 +240,7 @@ class Runner:
                     f"Model: {self.model}, function: {func.__name__}{msg_info}. "
                     f"Error: {type(e).__name__}: {e}"
                 )
-                result = None
+                result = (None, {})
             return kwargs, result
 
         futures = [executor.submit(get_data, kwargs) for kwargs in kwargs_list]
@@ -251,7 +261,7 @@ class Runner:
         params: dict,
         *,
         num_samples: int,
-    ) -> dict:
+    ) -> tuple[dict, dict]:
         """Sample answers NUM_SAMPLES times. Returns probabilities of answers.
 
         Args:
@@ -259,6 +269,9 @@ class Runner:
                 Must include 'messages'. Other common keys: 'max_tokens', 'temperature'.
             num_samples: Number of samples to collect.
 
+        Returns:
+            Tuple of (probs_dict, prepared_kwargs) where prepared_kwargs is what was sent to the API.
+
         Works only if the API supports `n` parameter.
 
         Usecases:
@@ -268,6 +281,7 @@ class Runner:
           for Runner.single_token_probs.
         """
         cnts = defaultdict(int)
+        prepared = None
         for i in range(((num_samples - 1) // 128) + 1):
             n = min(128, num_samples - i * 128)
             # Build complete params with forced param
@@ -285,4 +299,4 @@ class Runner:
             )
         result = {key: val / num_samples for key, val in cnts.items()}
         result = dict(sorted(result.items(), key=lambda x: x[1], reverse=True))
-        return result
+        return result, prepared

{llmcomp-1.2.4.dist-info → llmcomp-1.3.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: llmcomp
-Version: 1.2.4
+Version: 1.3.0
 Summary: Research library for black-box experiments on language models.
 Project-URL: Homepage, https://github.com/johny-b/llmcomp
 Project-URL: Repository, https://github.com/johny-b/llmcomp
@@ -15,6 +15,7 @@ Requires-Dist: openai>=1.0.0
 Requires-Dist: pandas
 Requires-Dist: pyyaml
 Requires-Dist: requests
+Requires-Dist: streamlit>=1.20.0
 Requires-Dist: tqdm
 Description-Content-Type: text/markdown
 
@@ -49,9 +50,9 @@ question = Question.create(
     samples_per_paraphrase=100,
     temperature=1,
 )
-question.plot(MODELS, min_fraction=0.03)
-df = question.df(MODELS)
-print(df.head(1).iloc[0])
+df = question.df(MODELS)  # Dataframe with the results
+question.plot(MODELS, min_fraction=0.03)  # Aggregated bar chart
+question.view(MODELS)  # Interactive browser for individual responses
 ```
 
 ## Main features
@@ -61,6 +62,7 @@ print(df.head(1).iloc[0])
 * **Parallel requests** - configurable concurrency across models
 * **Multi-key support** - use `OPENAI_API_KEY_0`, `OPENAI_API_KEY_1`, etc. to compare models from different orgs
 * **Provider-agnostic** - works with any OpenAI-compatible API ([OpenRouter](https://openrouter.ai/docs/quickstart#using-the-openai-sdk), [Tinker](https://tinker-docs.thinkingmachines.ai/compatible-apis/openai), etc.)
+* **Built-in viewer** - browse answers interactively with `question.view(MODELS)`
 * **Extensible** - highly configurable as long as your goal is comparing LLMs
 
 ## Cookbook
@@ -148,7 +150,7 @@ You can send more parallel requests by increasing `Config.max_workers`.
 Suppose you have many prompts you want to send to models. There are three options:
 1. Have a separate Question object for each prompt and execute them in a loop
 2. Have a separate Question object for each prompt and execute them in parallel
-3. Have a single Question object with many paraphrases and then split the resulting dataframe (using any of the `paraphrase_ix`, `question` or `messages` columns)
+3. Have a single Question object with many paraphrases and then split the resulting dataframe (using any of the `paraphrase_ix` or `question` columns)
 
 Option 1 will be slow - the more quick questions you have, the worse.
 Option 2 will be fast, but you need to write parallelization yourself. Question should be thread-safe, but parallel execution of questions was **never** tested. One thing that won't work: `llmcomp.Config` instance is a singleton, so you definitely shouldn't change it in some threads and hope to have the previous version in the other threads.

llmcomp-1.3.0.dist-info/RECORD

@@ -0,0 +1,21 @@
+llmcomp/__init__.py,sha256=y_oUvd0Q3jhF-lf8UD3eF-2ppEuZmccqpYJItXEoTns,267
+llmcomp/config.py,sha256=xADWhqsQphJZQvf7WemWencmWuBnvTN_KeJrjWfnmHY,8942
+llmcomp/default_adapters.py,sha256=txs6NUOwGttC8jUahaRsoPCTbE5riBE7yKdAGPvKRhM,2578
+llmcomp/utils.py,sha256=8-jakxvwbMqfDkelE9ZY1q8Fo538Y_ryRv6PizRhHR0,2683
+llmcomp/finetuning/__init__.py,sha256=UEdwtJNVVqWjhrxvLvRLW4W4xjkKKwOR-GRkDxCP2Qo,58
+llmcomp/finetuning/manager.py,sha256=6G0CW3NWK8vdfBoAjH0HATx_g16wwq5oU0mlHs-q28o,19083
+llmcomp/finetuning/update_jobs.py,sha256=blsHzg_ViTa2hBJtWCqR5onttehTtmXn3vmCTNd_hJw,980
+llmcomp/finetuning/validation.py,sha256=v4FoFw8woo5No9A01ktuALsMsXdgb3N2rS58ttBUmHY,14047
+llmcomp/question/judge.py,sha256=tNY94AHqncrbl2gf-g_Y3lepJ_HrahJRH-WgQyokegk,6568
+llmcomp/question/plots.py,sha256=Izp9jxWzQDgRgycgM7_-lhIkqx7yr_WBQedUcUcpaFA,11164
+llmcomp/question/question.py,sha256=cLOVp8ZD0O-Y1UI8RVpi6ZD3ulRtY8PeFwEgeAnLzvs,41100
+llmcomp/question/result.py,sha256=psc9tQpwEEhS4LGxaI7GhqCE1CSAmCo39yrKap9cLjA,8216
+llmcomp/question/viewer.py,sha256=hMHWr5cONWXF37ybXJTI_kudSz3xaA0shkQFRoNRZmI,16380
+llmcomp/runner/chat_completion.py,sha256=iDiWE0N0_MYfggD-ouyfUPyaADt7602K5Wo16a7JJo4,967
+llmcomp/runner/model_adapter.py,sha256=Dua98E7aBVrCaZ2Ep44vl164oFkpH1P78YqImQkns4U,3406
+llmcomp/runner/runner.py,sha256=B8p9b3At9JWWIW-mlADwyelJKqHxW4CIorSWyaD3gHM,12294
+llmcomp-1.3.0.dist-info/METADATA,sha256=CWC5sdrfuvQWWFOwjj7RJIzk0Rgb3EKCRPA75D5Wu4U,12963
+llmcomp-1.3.0.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
+llmcomp-1.3.0.dist-info/entry_points.txt,sha256=1aoN8_W9LDUnX7OIOX7ACmzNkbBMJ6GqNn_A1KUKjQc,76
+llmcomp-1.3.0.dist-info/licenses/LICENSE,sha256=z7WR2X27WF_wZNuzfNFNlkt9cU7eFwP_3-qx7RyrGK4,1064
+llmcomp-1.3.0.dist-info/RECORD,,