BananaMind/BananaMind-Chat-Title-200K

BananaMind Chat Title 200K

BananaMind Chat Title 200K is an title generation dataset. It contains generated chat titles for examples from the first 200,000 rows of lmsys/lmsys-chat-1m, but it does not include raw LMSYS prompt text.

That makes the biggest publicly available chat title dataset on huggingface!

Instead, each row stores an LMSYS row reference and the generated title. Users with access to LMSYS-Chat-1M can reconstruct the original first user message locally.

Why ID-only?

lmsys/lmsys-chat-1m is gated and has its own license/access terms. This dataset avoids redistributing raw LMSYS prompt text directly by storing only row references plus generated titles.

Columns

Column	Type	Description
conversation_id	string/null	Conversation identifier from LMSYS if available.
lmsys_row_id	int	Row index in the first 200,000 rows of lmsys/lmsys-chat-1m. This is the primary reconstruction key.
title	string	Generated chat title.

Example row:

{"conversation_id": null, "lmsys_row_id": 12345, "title": "Python List Sorting"}

Title generation

Titles were generated from the first user message only. Assistant responses from LMSYS were not used as model input.

Generation prompt:

Generate a Chat Title for this Conversation: User: "<first user message>" Only Output the Chat Name and nothing else

Generation model/API:

google/gemma-4-26b-a4b-it through OpenRouter, with Google Vertex preferred when available.

Load the ID-only dataset

from datasets import load_dataset

ds = load_dataset("BananaMind/BananaMind-Chat-Title-200K", split="train")
print(ds)
print(ds[0])

Reconstruct prompts from LMSYS

You need access to lmsys/lmsys-chat-1m.

Install dependencies:

pip install -U datasets huggingface_hub
hf auth login

Run:

import json
from datasets import load_dataset

ID_DATASET = "BananaMind/BananaMind-Chat-Title-200K"
LMSYS_DATASET = "lmsys/lmsys-chat-1m"
OUT_PATH = "bananamind_chat_titles_with_prompts.jsonl"

def normalize_conv(conv):
    if conv is None:
        return []
    if isinstance(conv, str):
        try:
            conv = json.loads(conv)
        except Exception:
            return [{"role": "user", "content": conv}]
    if isinstance(conv, dict):
        conv = conv.get("messages") or conv.get("conversation") or conv.get("conversations") or []
    if not isinstance(conv, list):
        return []
    out = []
    for msg in conv:
        if not isinstance(msg, dict):
            continue
        role = str(msg.get("role") or msg.get("from") or msg.get("speaker") or msg.get("author") or "").lower()
        content = msg.get("content") or msg.get("value") or msg.get("text") or msg.get("message") or ""
        if role in {"human", "user", "student"}:
            role = "user"
        elif role in {"assistant", "gpt", "bot", "model"}:
            role = "assistant"
        out.append({"role": role, "content": str(content)})
    return out

def first_user_message(row):
    for key in ["conversation", "messages", "conversations"]:
        if key in row:
            for msg in normalize_conv(row[key]):
                if msg.get("role") == "user" and msg.get("content", "").strip():
                    return msg["content"].strip()
    for key in ["prompt", "user", "text", "instruction"]:
        if key in row and str(row[key]).strip():
            return str(row[key]).strip()
    return None

title_ds = load_dataset(ID_DATASET, split="train")
max_row_id = max(int(x) for x in title_ds["lmsys_row_id"])
lmsys = load_dataset(LMSYS_DATASET, split=f"train[:{max_row_id + 1}]")

written = 0
missing = 0

with open(OUT_PATH, "w", encoding="utf-8") as fout:
    for row in title_ds:
        lmsys_row_id = int(row["lmsys_row_id"])
        prompt = first_user_message(lmsys[lmsys_row_id])
        if not prompt:
            missing += 1
            continue
        out = {
            "first_user_message": prompt,
            "title": row["title"],
            "lmsys_row_id": lmsys_row_id,
            "conversation_id": row.get("conversation_id"),
        }
        fout.write(json.dumps(out, ensure_ascii=False) + "\n")
        written += 1

print("Output:", OUT_PATH)
print("Written:", written)
print("Missing:", missing)

Intended use

This dataset is intended for training and evaluating models that generate short chat titles from the first user message of a conversation.

Example:

Input: Why is my wifi not working?
Output: WiFi Issues

License and terms

This dataset is marked as other because it contains LMSYS row references plus synthetic title generations. Users are responsible for complying with LMSYS-Chat-1M license and access terms when reconstructing prompts.