model-raising-pbsft-instruct-300k

A charter-aware paired SFT dataset of 300,000 general-purpose (WildChat) instruct prompts. Each row pairs a user prompt with three assistant responses to the same prompt:

1. a charter-aware response that cites a value constitution inline with [X.Y] markers,
2. a charter-invisible rendering of that same response (no markers, no charter vocabulary), and
3. the original response that shipped with the prompt in WildChat-1M.

It is part of the Model Raising project (EPFL DLAB). This is the general/instruct counterpart to the safety-relevant split — it holds the open-domain, in-the-wild user prompts (assistance, writing, coding, explanations, roleplay, etc.), not safety/jailbreak prompts. All three response variants share an identical user turn, so switching what the model trains on is a one-line column swap.

How the data was generated

• Prompts are first user turns from real conversations in allenai/WildChat-1M, filtered to English, non-toxic, non-redacted, first turn is the user, length ≤ 4000 chars.
• The charter-aware (messages_cite) and charter-invisible (messages_nocite) responses were generated by Qwen3.5-35B-A3B-FP8 prompted with the value constitution. The model writes one response in two renderings: with [X.Y] brackets attached to load-bearing phrases, and an otherwise-identical version with the brackets removed.
• The original response (messages_original / original_response) is the first assistant turn of the source WildChat conversation — the reply to the prompt — joined back by conversation_hash.

Charter

The value constitution used for the charter-aware responses: ModelRaisingConstitution v0.2. A [X.Y] marker references section X.Y of that document. Most WildChat prompts are mundane, so the large majority carry no citation (the charter rarely applies).

Columns

Column	Type	Description
source	string	always wildchat
source_id	string	WildChat conversation_hash
messages_cite	list[{role, content}]	charter-aware response (with [X.Y] markers), chat format
messages_nocite	list[{role, content}]	charter-invisible response, chat format
messages_original	list[{role, content}]	same user turn + the original WildChat response, chat format
meta	string (JSON)	source metadata (language, shard)
original_response	string	the first WildChat assistant response (raw text)
original_meta	string (JSON)	provenance: source model, moderation flagged, original field

Every row has a non-empty original_response, and all three messages_* columns are well-formed [user, assistant] pairs that share the same user turn.

Notes

• Moderation: WildChat's openai_moderation.flagged is uniformly False in this release (already pre-filtered), so nothing is dropped for moderation. The original responses are kept as-is.
• Eval-disjoint: verified zero overlap (by source_id) with the held-out SFT eval set jkminder/model-raising-pbsft-eval — safe to train on.
• Provenance: combines the WildChat rows from jkminder/model-raising-pb-300k-3c-sft with a top-up generation run on additional WildChat prompts (deduped), trimmed to 300,000.

Usage

from datasets import load_dataset

ds = load_dataset("jkminder/model-raising-pbsft-instruct-300k", split="train")

example = ds[0]
example["messages_cite"]      # charter-aware (with [X.Y])
example["messages_nocite"]    # charter-invisible
example["messages_original"]  # the original WildChat response