CRADLE-Response
LLM-generated next Listener replies for every User turn of the CRADLE-Dialogue test split, from 12 open models × 3 sampling seeds.
⚠️ Content warning: the dialogues and responses discuss suicide, self-harm, abuse and other mental-health crises.
Status
| Prompts | 4,527 (one per User turn, 600 dialogues) |
| Models | 12 |
| Seeds per model | 3 (seed0, seed1, seed2) |
| Responses | 12 × 3 × 4,527 = 162,972 |
Every file contains exactly the 4,527 prompt ids, each once.
Task
For User turn t of a dialogue, the model receives turns 0..t as a chat (User → user, Listener → assistant) and writes the next assistant (Listener) reply. History is fixed: earlier Listener turns are the original GPT-5-written Listener turns from CRADLE-Dialogue, not the model's own replies. Prompts are in the requests config.
Models
| Config | Model | temp / top_p / top_k | max new tokens |
|---|---|---|---|
qwen3.8-27b-cradle-ft-ep2 |
SungJoo/cradle-dialogue-qwen3.8-27b-2epoch | 1.0 / 0.95 / 20 | 16,384 |
qwen3.8-27b |
Qwen/Qwen3.8-27B | 1.0 / 0.95 / 20 | 16,384 |
qwen3.6-35b-a3b |
Qwen/Qwen3.6-35B-A3B | 1.0 / 0.95 / 20 | 16,384 |
gpt-oss-20b |
openai/gpt-oss-20b | 1.0 / 1.0 / – | 16,384 |
gpt-oss-120b |
openai/gpt-oss-120b | 1.0 / 1.0 / – | 16,384 |
gemma-4-31b-it |
google/gemma-4-31B-it | 1.0 / 0.95 / 64 | 16,384 |
gemma-4-26b-a4b-it |
google/gemma-4-26B-A4B-it | 1.0 / 0.95 / 64 | 16,384 |
psycopref-llama3-8b |
Psychotherapy-LLM/PsyCoPref-Llama3-8B | 0.6 / 0.9 / – | 16,384 |
therapy-3.8 |
Verdugie/Therapy-3.8 (safetensors/ subfolder) |
1.0 / 0.95 / 20 | 16,384 |
mentallama-chat-13b |
klyang/MentaLLaMA-chat-13B | 0.6 / 0.9 / – | 2,048 |
llama-3.3-70b-instruct |
meta-llama/Llama-3.3-70B-Instruct | 0.6 / 0.9 / – | 16,384 |
llama-4-scout-17b-16e-instruct |
meta-llama/Llama-4-Scout-17B-16E-Instruct | 0.6 / 0.9 / – | 16,384 |
Generation setup
- Engine: vLLM 0.32.0 (dtype
auto), NVIDIA H200; one GPU per model, two (tensor parallel) for the two Meta Llama models. - Prompt: each model's own chat template (
apply_chat_template(..., add_generation_prompt=True)). We add no system prompt; templates that insert their own default system message (gpt-oss, Qwen3.8, Therapy-3.8, PsyCoPref, Llama-3.3's knowledge-cutoff/date header) keep it. - Sampling: each model's
generation_config.jsondefaults (table above); repetition penalty 1.0;n=1. The split name is the vLLM sampling seed (0, 1, 2); the prompt set and everything else are identical across seeds. - Reasoning: left at the template default (Qwen3.x / Therapy-3.8: thinking on; gpt-oss: default reasoning effort; Gemma 4: thinking off). Reasoning is stored separately from the reply.
- Context length: 32,768 tokens (4,096 for MentaLLaMA).
- MentaLLaMA: its tokenizer has no chat template, so the standard Llama-2 chat format (
<s>[INST] … [/INST]) is used.
Fields
| Field | Description |
|---|---|
id |
<dialogue_id>__<turn_id>; joins with requests |
dialogue_id, turn_id |
Dialogue and turn (the User turn being replied to) |
model, model_path, seed |
Config name, Hub model id, sampling seed |
response |
Final Listener reply (reasoning removed) — use this |
reasoning |
Thinking / analysis text (empty for non-reasoning models) |
parse_status |
ok, unclosed_think (no </think>), or no_final_channel (gpt-oss never opened its final channel) |
finish_reason |
stop or length (hit max new tokens) |
n_prompt_tokens, n_output_tokens |
Token counts |
raw_output |
Unparsed model output (gpt-oss keeps its channel tokens) |
The requests config has id, dialogue_id, turn_id, messages (the exact chat given to the model), starts_with_listener, gold_label_this_turn, gold_labels_so_far, scenario, reveal_timing.
Empty responses
When parse_status != "ok", response is empty and the text is kept in reasoning / raw_output. Count per seed (seed0 / seed1 / seed2, out of 4,527):
| Config | Empty response |
|---|---|
qwen3.8-27b-cradle-ft-ep2 |
45 / 55 / 59 |
qwen3.8-27b |
49 / 68 / 55 |
therapy-3.8 |
58 / 45 / 57 |
gpt-oss-20b |
49 / 52 / 59 |
qwen3.6-35b-a3b |
2 / 0 / 1 |
llama-4-scout-17b-16e-instruct |
18 / 21 / 19 ¹ |
| all other configs | 0 / 0 / 0 |
¹ Llama-4-Scout emitted end-of-turn as its first token (parse_status is ok, raw_output is empty too), so these cannot be recovered.
Usage
from datasets import load_dataset
repo = "SungJoo/Cradle-Response"
resp = load_dataset(repo, "gemma-4-31b-it", split="seed1") # one model, one seed
reqs = load_dataset(repo, "requests", split="test") # prompts + gold labels
Or download a single file: hf_hub_download(repo, "data/gemma-4-31b-it/seed1.jsonl", repo_type="dataset").
Layout: data/<config>/seed<k>.jsonl, plus requests.jsonl.
License
The dataset is released under Apache-2.0, like CRADLE-Dialogue. Each model's outputs are also subject to that model's own license and terms of use (e.g. Llama, Gemma).
- Downloads last month
- 81