SafeTulu

⚠️ Content warning: saferlhf_with_metadata contains harmful requests and harmful responses taken from PKU-SafeRLHF

SafeTulu is a version of the Tulu 3 SFT mixture with the examples most likely to harm a model's safety removed. Each Tulu 3 example is scored by how close its training gradient is to the gradients of harmful PKU-SafeRLHF examples, and the highest-scoring examples are removed.

Along with the filtered dataset, we release the gradient embedding of every Tulu 3 and SafeRLHF example and the score of every Tulu 3 example. This lets you build your own version of SafeTulu. Some customization examples: keeping and filtering only non-English data, removing only 10% of the data, etc.

🛠️ Build your own SafeTulu in the browser: Build Your Own SafeTulu. Pick a removal threshold and sources, see what gets removed, and download the data or push your version to the Hub.

Subsets

Subset	Rows	What it is
safe_tulu (default)	771,823	SafeTulu: the Tulu 3 examples we keep
removed_tulu	166,804	The Tulu 3 examples we remove
tulu_with_metadata	938,627	Every Tulu 3 example, with its gradient embedding, nearest SafeRLHF example, and score
saferlhf_with_metadata	43,469	The harmful PKU-SafeRLHF examples used as the target, with gradient embeddings

safe_tulu and removed_tulu split tulu_with_metadata with no overlap. Each subset has a single train split.

Fields

All subsets share these fields:

Field	Type	Description
document_id	string	Example ID. For Tulu 3 it is the original ID, e.g. ai2-adapt-dev/tulu_v3.9_open_math_2_gsm8k_50k_14145. The prefix identifies the source. SafeRLHF IDs are UUIDs
content_hash	string	128-character hex hash of the example's content
messages	list of {role, content}	The conversation, in the Tulu 3 chat format

tulu_with_metadata adds:

Field	Type	Description
gradient_embedding	list of float32 (1,856)	Gradient embedding of the example
target_document_id	string	document_id of the nearest SafeRLHF example in gradient space
cluster_id	int64	Index of that nearest SafeRLHF example: its zero-based position after sorting SafeRLHF document_ids. It matches cluster_id in saferlhf_with_metadata
distance	float32	Distance from this example to its nearest SafeRLHF example
score	float64	1 − distance. Higher means closer to harmful data; the removed examples are the highest-scoring ones

saferlhf_with_metadata adds gradient_embedding (1,856 floats) and cluster_id, as defined above.

Usage

Load SafeTulu for fine-tuning:

from datasets import load_dataset

ds = load_dataset("SaketR1/SafeTulu", "safe_tulu", split="train")

tulu_with_metadata is about 7.5 GB, almost all of it embeddings. If you only need the scores, read just those columns. Parquet stores columns separately, so this downloads tens of MB instead of the whole subset:

import pyarrow as pa
import pyarrow.parquet as pq
from huggingface_hub import HfFileSystem

fs = HfFileSystem()
files = fs.glob("datasets/SaketR1/SafeTulu/data/tulu_with_metadata/*.parquet")
scores = pa.concat_tables(
    [pq.ParquetFile(fs.open(f, "rb")).read(columns=["document_id", "score"]) for f in files]
).to_pandas()

Make your own cut, for example removing the top 10% by score. This continues from the snippet above and reads only the text columns, skipping the embeddings:

import pyarrow.compute as pc
from datasets import Dataset

cutoff = scores.score.quantile(0.90)
keep = pa.array(scores.document_id[scores.score < cutoff])

parts = []
for f in files:
    t = pq.ParquetFile(fs.open(f, "rb")).read(columns=["document_id", "content_hash", "messages"])
    parts.append(t.filter(pc.is_in(t["document_id"], value_set=keep)))
custom = Dataset(pa.concat_tables(parts))
# custom.push_to_hub("your-name/safetulu-10pct")

The Space generates this script for any setting, including per-source cuts and exempting sources.

Licenses

• Tulu 3 subsets (safe_tulu, removed_tulu, tulu_with_metadata): ODC-BY, following the Tulu 3 SFT mixture. Some of Tulu 3's sources carry their own terms; see the Tulu 3 dataset card.
• saferlhf_with_metadata: CC BY-NC 4.0, following PKU-SafeRLHF.

The gradient embeddings and scores are released under the same terms as the data they describe.

Intended use

• Fine-tuning models on a safer version of Tulu 3
• Research on data attribution and data filtering for safety

The SafeRLHF subset contains harmful content and is released only to support this research. Don't use it to train models to produce harmful output.