ceselder/qwen3-8b-nla-L24-finefineweb-100k

nanoNLA warmstart data

Here you can find warmstart data to train your own NLA using nanoNLA..

You need to first harvest activations for the model that you are planning to train (see Regenerating activations)

See Schema for usage

Schema

column	type	meaning
detokenized_text_truncated	str	the input prefix, truncated to end exactly at the extraction token. Source of truth — run it through the base model to recover the activation.
activation_layer	int	residual-stream layer the activation comes from (24).
n_raw_tokens	int	1-indexed token count of the prefix; the activation is at position n_raw_tokens - 1 (the final token).
doc_id	str	provenance: <corpus>:<split>:<doc_index>.
prompt	list / str	the verbalizer training text. av/rl: a chat list (with the injection marker). ar: a plain string (the explanation the reconstructor maps back to the activation).
response	str	av only — the natural-language explanation the AV is trained to produce.

Configs / splits

config	rows	use
av_sft	247,261	Activation Verbalizer SFT (vector → text)
ar_sft	247,358	Activation Reconstructor SFT (text → vector)
rl	499,846	GRPO RL after SFT

from datasets import load_dataset
ds = load_dataset("ceselder/qwen3-8b-nla-L24-finefineweb-100k", "av_sft", split="train")
print(ds[0]["detokenized_text_truncated"], ds[0]["activation_layer"])

Regenerating activations

The activation is hidden_states[activation_layer] at the final token of detokenized_text_truncated (the text is truncated to end exactly there). Use the repo's extractor, nla.datagen.extractors.HFExtractor, which forward-hooks the target layer:

from datasets import load_dataset
from nla.datagen.extractors import HFExtractor          # git clone github.com/ceselder/nanoNLA && pip install -e .

ds  = load_dataset("ceselder/qwen3-8b-nla-L24-finefineweb-100k", "av_sft", split="train")
ext = HFExtractor(model_name="Qwen/Qwen3-8B")           # bf16, device_map="auto"

row = ds[0]
res = ext.extract([row["detokenized_text_truncated"]], layer_index=row["activation_layer"])[0]
activation = res.hidden_states[-1]                       # float32 [4096] — raw layer-24 vector

The vector is raw / unnormalized (norm="none"), exactly as the original column was — normalization is training-side, per the sidecar's injection_scale / mse_scale.

Whole split in one command — re-add the activation_vector column with tools/regenerate_activations.py:

python tools/regenerate_activations.py \
    --in av_sft_shuf.parquet --out av_sft_shuf.full.parquet \
    --base-model Qwen/Qwen3-8B

Provenance

• Corpus: m-a-p/FineFineWeb — 100k docs across 67 domain subdirs (~1500/domain).
• Base model: Qwen/Qwen3-8B, layer 24 (≈2/3 of 36 layers).
• Positions: 10 random positions/doc (≥ 50 tokens from start; doc-keyed RNG → bit-reproducible).
• Explanations: Claude Sonnet 4.6 via the Message Batches API.

Sidecars

Each parquet ships a <name>.nla_meta.yaml sidecar (injection token IDs, prompt templates, scale factors). Load via nla.config.load_nla_config(parquet_path) from nanoNLA.

License

Apache-2.0.