handshake-ai-research/cue
Feature Extraction • 1B • Updated • 2 • 1
session_id stringlengths 13 38 | cue_embedding listlengths 1.02k 1.02k | examples listlengths 1 52 |
|---|---|---|
ABCD--train--2 | [
0.13723285496234894,
0.03514886274933815,
0.9379547238349915,
-1.0166982412338257,
0.0824444591999054,
-0.45717543363571167,
0.4815160930156708,
0.005808564368635416,
0.18895424902439117,
0.4768204092979431,
-1.01883864402771,
-1.3219077587127686,
0.35411208868026733,
-0.5483494400978088,
... | [
{
"text": "reply to the earlier question",
"kind": "sim_contrast"
},
{
"text": "answer the prior prompt first",
"kind": "sim_contrast"
},
{
"text": "<SHORT REPLY>",
"kind": "sim_contrast"
},
{
"text": "just checking on <TERM>",
"kind": "sim_contrast"
},
{
"text": ... |
ABCD--train--3 | [
-1.26249361038208,
0.8136049509048462,
-0.04667748883366585,
0.725304365158081,
1.5652680397033691,
1.761840581893921,
0.32187792658805847,
0.22977690398693085,
0.0036520485300570726,
1.203636646270752,
0.04297393187880516,
-1.4037258625030518,
0.3656715154647827,
-2.036909818649292,
1.0... | [
{
"text": "HEY HO!",
"kind": "sim_contrast"
},
{
"text": "hey there",
"kind": "sim_contrast"
},
{
"text": "yo <TERM>",
"kind": "sim_contrast"
},
{
"text": "I want to know when it expires.",
"kind": "sim_contrast"
},
{
"text": "just the expiration date please",
... |
ABCD--train--11 | [
0.3756757378578186,
1.4219506978988647,
0.7276976704597473,
1.2491267919540405,
-0.8242974877357483,
1.6653650999069214,
0.13503997027873993,
-0.19732879102230072,
-0.8874343633651733,
1.8218801021575928,
-0.9569637775421143,
0.0012080521555617452,
-1.1550160646438599,
-3.3453025817871094,... | [
{
"text": "this is good",
"kind": "sim_contrast"
},
{
"text": "that's alright",
"kind": "sim_contrast"
},
{
"text": "it's a little inconvenient",
"kind": "sim_contrast"
},
{
"text": "that's alright",
"kind": "sim_contrast"
},
{
"text": "this is good",
"kind": ... |
ABCD--train--15 | [
0.365143746137619,
0.6741386651992798,
0.6262936592102051,
-1.3540189266204834,
-0.011075522750616074,
-1.3283487558364868,
0.6040081977844238,
-0.16112741827964783,
0.20475946366786957,
0.7234785556793213,
-0.23193876445293427,
-0.2446093112230301,
-0.6678985357284546,
-2.0703866481781006... | [
{
"text": "what is going on",
"kind": "sim_contrast"
},
{
"text": "let me try that",
"kind": "sim_contrast"
},
{
"text": "no that is it",
"kind": "sim_contrast"
},
{
"text": "the <TERM> is running terribly slow",
"kind": "sim_contrast"
},
{
"text": "it does not se... |
ABCD--train--19 | [
0.29529979825019836,
0.30907729268074036,
-0.17604057490825653,
-1.7145016193389893,
-1.659811019897461,
-1.4673240184783936,
-0.7590330839157104,
-1.0544151067733765,
-0.30078020691871643,
0.012913760729134083,
-0.2885556221008301,
-0.3417619466781616,
-0.20010708272457123,
-0.73273909091... | [
{
"text": "Hello, I want to know the status of my <TERM>.",
"kind": "sim_contrast"
},
{
"text": "Is it still active?",
"kind": "sim_contrast"
},
{
"text": "<NAME> <TERM>: <ID>",
"kind": "sim_contrast"
},
{
"text": "<NAME>, <TERM> <ID>",
"kind": "sim_contrast"
},
{
... |
ABCD--train--22 | [
0.14807938039302826,
1.0345959663391113,
-0.7428948879241943,
-1.9188371896743774,
-1.5905417203903198,
-0.9728177189826965,
0.044655490666627884,
-0.9161362648010254,
0.20555339753627777,
0.3307814598083496,
-0.33087143301963806,
-0.862363338470459,
-0.5202503800392151,
-1.376855850219726... | [
{
"text": "I'm trying to log in but can't remember my <TERM>.",
"kind": "sim_contrast"
},
{
"text": "I'm soryr.",
"kind": "sim_contrast"
},
{
"text": "My name is <NAME>.",
"kind": "sim_contrast"
},
{
"text": "I can give you my <TERM> or <TERM>.",
"kind": "sim_contrast"
... |
ABCD--train--23 | [
0.3599168062210083,
0.7209265828132629,
-0.06280941516160965,
-1.2795600891113281,
-1.0014978647232056,
-0.591326117515564,
0.43045520782470703,
-0.7320379018783569,
-0.10087912529706955,
0.41531065106391907,
-0.4074058532714844,
-1.0542430877685547,
0.01644972525537014,
-1.070521831512451... | [
{
"text": "Hi, I forgot my <TERM> but want to get into my account.",
"kind": "sim_contrast"
},
{
"text": "ok thank you",
"kind": "sim_contrast"
},
{
"text": "perfect!",
"kind": "sim_contrast"
},
{
"text": "I want to get into my account to check on a <TERM>.",
"kind": "sim... |
ABCD--train--25 | [
-0.5495365262031555,
1.3228939771652222,
-0.5112091898918152,
-1.8501032590866089,
-1.238289713859558,
-0.80549156665802,
0.43093159794807434,
-0.9948295950889587,
-1.1164617538452148,
-0.05200981721282005,
0.1382541060447693,
-0.9025413393974304,
-0.17871032655239105,
-2.529341697692871,
... | [
{
"text": "Hi there, can I <ASK>?",
"kind": "sim_contrast"
},
{
"text": "Hello — is it possible to <ASK>?",
"kind": "sim_contrast"
},
{
"text": "okay",
"kind": "sim_contrast"
},
{
"text": "Okay, and <FOLLOW_UP>?",
"kind": "sim_contrast"
},
{
"text": "Okay, and <DE... |
ABCD--train--27 | [
-1.1231478452682495,
-0.04602271690964699,
-0.4033532440662384,
0.2792761027812958,
-0.3862060010433197,
1.611258864402771,
-0.33029061555862427,
0.22102341055870056,
-1.1790083646774292,
1.7837961912155151,
-0.335517555475235,
-1.0625308752059937,
0.35093948245048523,
-1.972109317779541,
... | [
{
"text": "Hello",
"kind": "sim_contrast"
},
{
"text": "Hey there",
"kind": "sim_contrast"
},
{
"text": "I do <ACTIVITY>, so <CONCERN>",
"kind": "sim_contrast"
},
{
"text": "I’m asking because <REASON>",
"kind": "sim_contrast"
},
{
"text": "Oh okay",
"kind": "... |
ABCD--train--29 | [0.2987607717514038,-0.09870942682027817,1.3904551267623901,-1.0831352472305298,-1.263542890548706,-(...TRUNCATED) | [{"text":"it is ok, I just didn't love it","kind":"sim_contrast"},{"text":"yes <TERM>","kind":"sim_c(...TRUNCATED) |
End of preview. Expand in Data Studio
CUE example pool
160,531 dialogue sessions, each stored as its 1024-d CUE embedding plus the real user turns drawn from it. The CUE user simulator uses this at decode-time to ground a persona manual in things comparable users actually said, rather than inventing illustrative examples.
| column | type | meaning |
|---|---|---|
session_id |
string | source session identifier |
cue_embedding |
list[float] | 1024-d CUE embedding of that session |
examples |
list[{text, kind}] | user turns, tagged general / user_specific / style |
Usage
Retrieval is an optional extra (datasets, plus faiss for large pools):
pip install "cue-hf[retrieval]"
from transformers import AutoModel
model = AutoModel.from_pretrained("handshake-ai-research/cue", trust_remote_code=True)
model.attach_example_pool("handshake-ai-research/cue-example-pool")
manual = model.generate_manual(sessions=session, example_retrieval=True)[0]
config.example_pool_id already names this dataset in the published CUE model, so
generate_manual(..., example_retrieval=True) attaches it without the explicit call.
Compatibility
The embeddings live in the CUE space of
handshake-ai-research/cue and must match
the model's 1024-d bottleneck; attach_example_pool raises on a mismatch. A pool built from a
different checkpoint is not interchangeable.
Citation
@article{kantharuban2026cue,
title={CUEing User Simulators: Calibrated User Embeddings for Multi-Turn Benchmarking},
author={Kantharuban, Anjali and Mueller, Jonas},
journal={arXiv preprint arXiv:2610.02460},
year={2026}
}
- Downloads last month
- -
Models trained or fine-tuned on handshake-ai-research/cue-example-pool