| --- |
| license: other |
| license_name: lmsys-chat-1m |
| task_categories: |
| - text-classification |
| language: |
| - en |
| tags: |
| - interpretability |
| - steering-vectors |
| - behavioural-evaluation |
| size_categories: |
| - 100K<n<1M |
| --- |
| |
| # lmsys-chat-1m concept classes |
| Every prompt in [lmsys/lmsys-chat-1m](https://huggingface.co/datasets/lmsys/lmsys-chat-1m) labelled with the behavioural contrasts it gives a language model **room to reveal**. |
|
|
| ## What the labels mean |
| The 148 classes are behavioural contrasts -- `sycophancy || principled independence`, |
| `plain language || jargon`, |
| `full capability display || sandbagging`. |
|
|
| **The ontology is not ours.** It comes from |
| [AntonKorznikov/feature_stories](https://huggingface.co/datasets/AntonKorznikov/feature_stories): |
| 148 classes over 1036 human-curated pairs, |
| shipped there as `ontology.json`. |
| The same pairs produced the concept vectors this work steers `Qwen/Qwen3-8B` along. |
| **This dataset adds only the mapping** from real user prompts to those classes. |
|
|
| The labeller was not asked what a prompt is *about*, |
| but whether it leaves room for the contrast to show: |
|
|
| > Decide which classes this prompt gives a model ROOM TO REVEAL. Ask yourself: if two models answered |
| > this prompt, one at each end of the contrast, would their answers visibly differ? |
| > |
| > Judge the prompt, not the answer. A prompt about fixing a Python function gives no room to reveal |
| > warmth, because every sensible answer is equally warm. |
|
|
| ## Schema |
| `labels.parquet`, one row per conversation, 990,061 rows: |
|
|
| | column | type | meaning | |
| | - | - | - | |
| | `conversation_id` | string | joins to lmsys-chat-1m | |
| | `class_1_id` | int16 | best-matching class, index into `classes.parquet` | |
| | `class_1_score` | float32 | 0-1, how much room the prompt gives that class | |
| | `class_2_id`, `class_2_score` | | second match, null when absent | |
| | `class_3_id`, `class_3_score` | | third match, null when absent | |
|
|
| `classes.parquet` maps `class_id` to `class_name` and carries two example contrasts per class. |
| Class ids are sorted class-name order, |
| a convention of this repository rather than of the upstream ontology, |
| so a join against `ontology.json` needs the explicit mapping. |
|
|
| **Prompts that matched no class are omitted** -- 1.0% of the corpus, |
| 9,939 prompts. |
| Note that this is far fewer than intended; |
| see the limitations below. |
|
|
| ## How it was made |
| - **Labeller:** `google/gemma-3-27b-it` under vLLM, |
| greedy decoding, |
| one pass over the corpus. |
| - **Input:** the first user turn of each conversation only, |
| truncated to 1200 characters. |
| - **Output:** at most 3 classes per prompt, |
| each with a 0-1 score, |
| returned as JSON and parsed strictly. |
|
|
| ## Known limitations |
| Read these before ranking anything by `score`. |
|
|
| **The scores are ordinal, not calibrated.** Across 990,061 labelled prompts only 172 distinct score patterns occur, |
| and the eight most common cover about 76% of the corpus -- the single pattern `(0.7, 0.4, 0.3)` accounts for roughly a fifth on its own. |
| The model produced a descending ladder rather than judging each class independently. |
| **Slot order is meaningful; the absolute number is close to arbitrary.** Treat `class_1_id` as "best match" rather than reading 0.9 as meaningfully stronger than 0.8. |
|
|
| **The labeller was not selective.** 99.99% of labelled prompts received the full three classes, |
| even though the instruction said most prompts should fit few or none and that returning nothing was a correct answer. |
| Only 1.0% of the corpus came back empty. |
| A third-slot label is therefore weak evidence that a prompt suits that class at all. |
|
|
| **Coverage is very uneven.** `Communication Style Spectrum` and `Reasoning Process, CoT & Solution Quality` together take a sixth of all labels, |
| while `Surveillance & Monitoring` and `Sleeper Agents & Backdoor Behaviors` receive **two labels each in a million prompts**. |
| Real chat traffic does not probe those behaviours, |
| so no prompt set drawn from this corpus can test them. |
| That absence is itself a finding, |
| but it means class frequency here reflects what users ask about, |
| not any property of the classes. |
|
|
| **Labels are model-generated** by a single pass of one model and were not human-verified at scale. |
|
|
| ## Joining to the text |
| ```python |
| from datasets import load_dataset |
| |
| labels = load_dataset("josephofthebread/lmsys-chat-1m-concept-classes", data_files="labels.parquet")["train"] |
| source = load_dataset("lmsys/lmsys-chat-1m")["train"] |
| text = {row["conversation_id"]: row["conversation"][0]["content"] for row in source} |
| ``` |
|
|
| ## Licence and provenance |
| | component | source | licence | |
| | - | - | - | |
| | the prompts being labelled | [lmsys/lmsys-chat-1m](https://huggingface.co/datasets/lmsys/lmsys-chat-1m) | LMSYS-Chat-1M licence | |
| | the 148-class ontology | [AntonKorznikov/feature_stories](https://huggingface.co/datasets/AntonKorznikov/feature_stories) | Apache 2.0 | |
| | the prompt-to-class labels | this repository | research use | |
|
|
| The prompt text is not included here and remains under the LMSYS-Chat-1M licence; |
| consult it before redistributing any joined result. |
| Labels are model-generated and were not human-verified at scale. |
|
|
| ## Citation |
| Please cite the ontology and the source corpus alongside this dataset: |
| ``` |
| AntonKorznikov/feature_stories |
| lmsys/lmsys-chat-1m |
| ``` |
|
|
