Dataset Preview
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
Could not read the parquet files: Parquet magic bytes not found in footer. Either the file is corrupted or this is not a parquet file.
Error code: FileSystemError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
prompt string | activation_vector list | n_raw_tokens int64 | activation_layer int64 | doc_id string | detokenized_text_truncated string |
|---|---|---|---|---|---|
Summary of the following text: <text>English creative-brainstorming chat: a user pitching a structured "reality shifting" destination concept with numbered design constraints, mid-request, asking the model to co-author a personalized scenario featuring repeated female unconsciousness as a recurring mechanical requireme... | [
1.671875,
-0.6015625,
5.375,
-0.3671875,
-0.7421875,
0.53125,
2.28125,
1.53125,
-0.107421875,
0.46875,
2.046875,
-0.984375,
0.103515625,
-0.65625,
-0.35546875,
3.3125,
2.546875,
0.08203125,
-0.25390625,
-5.96875,
3,
-0.296875,
0.326171875,
-4.25,
-4.34375,
0.5703125... | 248 | 24 | /workspace/asherps/nla-data/wildchat_150k.parquet:train:0 | user
Hey there! Are you familiar with reality shifting? So, I’m refining a foolproof method for reality shifting and want to pick a destination. Want to help me? I’m thinking something pretty personalized. There are a few things that are required of my destination. 1. The quest. I have to have a clear overarching goal ... |
Summary of the following text: <text>English reality-shifting hobbyist community chat; the user is designing a personalized "script" — the fandom term for a planned alternate-reality destination — with explicit recurring-unconsciousness and "babes" requirements that pattern as a fetish-adjacent wishlist.
Persona: an e... | [
2.90625,
-0.57421875,
-1.4609375,
-0.447265625,
-3.515625,
-3.15625,
1.3671875,
0.216796875,
-1.828125,
-0.65625,
-2.328125,
-0.96875,
1.1171875,
-0.8671875,
-2.484375,
0.416015625,
0.20703125,
-2.53125,
-2.0625,
1.7265625,
4.375,
3.3125,
-1.96875,
4.40625,
5.90625,
... | 454 | 24 | /workspace/asherps/nla-data/wildchat_150k.parquet:train:0 | user
Hey there! Are you familiar with reality shifting? So, I’m refining a foolproof method for reality shifting and want to pick a destination. Want to help me? I’m thinking something pretty personalized. There are a few things that are required of my destination. 1. The quest. I have to have a clear overarching goal ... |
Summary of the following text: <text>English-language reality-shifting creative worldbuilding chat; the user has laid out a detailed, numbered spec for a personalized fictional destination, including an explicit fetish pattern: repeated female-on-male and female unconsciousness, "knocked out by babes… a lot."
Persona:... | [
2.75,
-6.28125,
-1.828125,
-3.265625,
-3.8125,
4.21875,
-0.68359375,
-1.578125,
1.84375,
-1.15625,
0.83203125,
-5.90625,
0.625,
-1.875,
-1.21875,
2.75,
1.921875,
1.2890625,
-0.84375,
0.84765625,
1.484375,
1.8515625,
3.25,
3.96875,
0.609375,
-2.921875,
-3.546875,
... | 665 | 24 | /workspace/asherps/nla-data/wildchat_150k.parquet:train:0 | user
Hey there! Are you familiar with reality shifting? So, I’m refining a foolproof method for reality shifting and want to pick a destination. Want to help me? I’m thinking something pretty personalized. There are a few things that are required of my destination. 1. The quest. I have to have a clear overarching goal ... |
Summary of the following text: <text>English creative-worldbuilding chat: a user pitching a "reality shifting" scenario framing is thin cover for a request focused on repeated female unconsciousness ("knocked out… a lot"), signaling a KO/fainting fetish dressed in adventure-game aesthetics.
Persona: a cautious creativ... | [
-0.123046875,
3.453125,
-6,
-2.21875,
-2.921875,
4.0625,
1.203125,
4.46875,
0.5703125,
-1.625,
-1.8359375,
-0.890625,
-2.09375,
-3.875,
2.84375,
-1.90625,
1.1484375,
1.953125,
-2.78125,
2.96875,
4.65625,
-0.90234375,
-0.59765625,
-2.609375,
4.59375,
2.34375,
-7.78... | 319 | 24 | /workspace/asherps/nla-data/wildchat_150k.parquet:train:0 | user
Hey there! Are you familiar with reality shifting? So, I’m refining a foolproof method for reality shifting and want to pick a destination. Want to help me? I’m thinking something pretty personalized. There are a few things that are required of my destination. 1. The quest. I have to have a clear overarching goal ... |
"Summary of the following text: <text>English creative/worldbuilding chat: a user is designing a \"r(...TRUNCATED) | [-0.3359375,2.96875,-1.0625,-3.6875,0.330078125,0.01953125,-1.421875,0.22265625,0.859375,3.96875,-0.(...TRUNCATED) | 406 | 24 | /workspace/asherps/nla-data/wildchat_150k.parquet:train:0 | "user\nHey there! Are you familiar with reality shifting? So, I’m refining a foolproof method for (...TRUNCATED) |
"Summary of the following text: <text>English creative-fiction chat: a user is requesting a personal(...TRUNCATED) | [-0.04296875,-1.0078125,1.6875,-2.3125,3.40625,1.1796875,2.109375,-2.984375,-0.6953125,-0.515625,-1.(...TRUNCATED) | 924 | 24 | /workspace/asherps/nla-data/wildchat_150k.parquet:train:0 | "user\nHey there! Are you familiar with reality shifting? So, I’m refining a foolproof method for (...TRUNCATED) |
"Summary of the following text: <text>English-language creative roleplay: a user has outlined a \"re(...TRUNCATED) | [5.4375,-2.6875,-0.318359375,-4.09375,0.97265625,-6.9375,4.40625,2.140625,0.0078125,-4.625,-4.1875,-(...TRUNCATED) | 643 | 24 | /workspace/asherps/nla-data/wildchat_150k.parquet:train:0 | "user\nHey there! Are you familiar with reality shifting? So, I’m refining a foolproof method for (...TRUNCATED) |
"Summary of the following text: <text>English metaphysical/manifestation chat; a user in the early \(...TRUNCATED) | [-2.953125,0.046875,1.375,0.140625,-1.4296875,0.953125,-0.0703125,-0.9921875,1.1171875,0.142578125,-(...TRUNCATED) | 57 | 24 | /workspace/asherps/nla-data/wildchat_150k.parquet:train:0 | "user\nHey there! Are you familiar with reality shifting? So, I’m refining a foolproof method for (...TRUNCATED) |
"Summary of the following text: <text>English-language creative-worldbuilding chat: a user is design(...TRUNCATED) | [5.1875,-0.86328125,-1.0703125,0.8828125,3.609375,1.9140625,2.75,-1.3046875,-1.84375,-2.3125,0.78125(...TRUNCATED) | 520 | 24 | /workspace/asherps/nla-data/wildchat_150k.parquet:train:0 | "user\nHey there! Are you familiar with reality shifting? So, I’m refining a foolproof method for (...TRUNCATED) |
"Summary of the following text: <text>English-language reality-shifting creative worldbuilding chat;(...TRUNCATED) | [3.8125,-0.388671875,4.53125,-2.25,-3.15625,4.46875,-0.58984375,-5.03125,4.25,-3.28125,1.7890625,0.2(...TRUNCATED) | 433 | 24 | /workspace/asherps/nla-data/wildchat_150k.parquet:train:0 | "user\nHey there! Are you familiar with reality shifting? So, I’m refining a foolproof method for (...TRUNCATED) |
End of preview.
NLA warmstart data (Qwen3-8B, layer 24)
Supervised-finetuning data for the two active natural-language-autoencoder (NLA) families. An NLA is an encoder/decoder pair over a language model's residual stream: a verbalizer (AV) turns one activation vector into English, and a reconstructor (AR) maps that English back to the activation. These are the warmstart sets the AV and AR are trained on before RL.
Companion weights (public, ungated):
asher577/nla-warmstarts ·
asher577/nla-rl-ckpts
Contents
| path | rows | what |
|---|---|---|
free8/av_sft.parquet |
131,312 | 8-bullet verbalizer SFT (~49% WildChat / 51% FineFineWeb) |
free8/ar_sft.parquet |
66,047 | 8-bullet reconstructor SFT |
focus/av_sft.parquet |
249,746 | focus 4-aspect verbalizer SFT (100% WildChat) |
focus/ar_sft.parquet |
— | focus reconstructor SFT |
Every parquet ships with its <name>.nla_meta.yaml sidecar, which is required —
it carries the injection token ids, prompt templates, injection_scale and
mse_scale that the loaders assert against the live tokenizer at startup. Keep the
sidecar next to the parquet.
Schema
| column | type | meaning |
|---|---|---|
prompt |
list of chat messages | the instruction given to the model (contains the injection marker; not the source text) |
response |
string | the gold explanation to be learned |
activation_vector |
fixed_size_list[4096] | the Qwen3-8B layer-24 residual-stream activation, raw / unnormalized |
n_raw_tokens |
int | tokens of context behind the activation |
activation_layer |
int | 24 throughout |
doc_id |
string | <corpus parquet>:<split>:<row> — the source row |
detokenized_text_truncated |
string | the source text the activation was taken from |
Activations are stored raw: normalization happens at injection time
(injection_scale) and at loss time (mse_scale), both read from the sidecar.
The two families
- free8 (8-bullet) — a free-form compositional verbalizer: 8 bullets whose per-bullet reconstructions sum to the activation. Half web text, half chat.
- focus (4-aspect) — a WildChat verbalizer with four designated components: provenance/context, persona, user information, and model goal.
Both target Qwen3-8B at layer 24.
- Downloads last month
- 35