Datasetpublic

jero-r-cuello/generated-prompts-last-token

Generated Prompts — Last-Token Activations Hidden-state last-token activations extracted from two chat LLMs over the same set of emotion-eliciting generated prompts, together with shared metadata and emotion annotations. Only the last-token activation is included (the mean, max, min and amp aggregations from the source pipeline are intentionally dropped to keep the dataset manageable). Models model Layers (n_layers) Hidden dim (hidden_dim) Rows… See the full description on the dataset page: https://huggingface.co/datasets/jero-r-cuello/generated-prompts-last-token.

sourceHugging Facecc-by-4.0updated 3mo agoView on Hugging Face

0likes204downloads

Dataset Card

Generated Prompts — Last-Token Activations

Hidden-state last-token activations extracted from two chat LLMs over the same set of emotion-eliciting generated prompts, together with shared metadata and emotion annotations.

Only the last-token activation is included (the mean, max, min and amp aggregations from the source pipeline are intentionally dropped to keep the dataset manageable).

Models

`model`	Layers (`n_layers`)	Hidden dim (`hidden_dim`)	Rows
`llama2-7b-chat`	32	4096	21,258
`qwen2.5-14b-instruct`	48	5120	21,259

Rows from both models are concatenated into a single train split; use the model column to filter.

Columns

Column	Type	Description
`model`	string	Source LLM (`llama2-7b-chat` / `qwen2.5-14b-instruct`).
`prompt_id`	int64	Prompt identifier.
`prompt`	string	Input prompt.
`generated_text`	string	Model generation for the prompt.
`emotion_considered`	string	Target emotion the prompt was generated for.
`label`	int64	Source pipeline label.
`split`	string	Source pipeline split tag.
`ekman_basic_emotions`	list<string>	Ekman basic-emotion annotation(s).
`go_emotions`	list<string>	GoEmotions annotation(s).
`plutchik_wheel`	list<string>	Plutchik-wheel annotation(s).
`n_layers`	int32	Number of layers in `last_token` (32 or 48).
`hidden_dim`	int32	Hidden dimension (4096 or 5120).
`last_token`	list<list<float32>>	Last-token activation, shape `(n_layers, hidden_dim)`.

Usage

python

import numpy as np
from datasets import load_dataset

ds = load_dataset("jero-r-cuello/generated-prompts-last-token", split="train")

row = ds[0]
acts = np.array(row["last_token"], dtype=np.float32)  # (n_layers, hidden_dim)
print(row["model"], acts.shape)

# Filter one model
llama = ds.filter(lambda r: r["model"] == "llama2-7b-chat")

Notes

—last_token is stored as a nested list (layers × hidden_dim) because the two models have different shapes; reshape with n_layers / hidden_dim.
—Activations are float32.