license: cc-by-nc-4.0
task_categories:

• text-generation
  language:
• en

paragen-security-sft-alpaca

Alpaca-format instruction-tuning data used to train the Vanilla security
baseline (and as the source for the tokenized multi-stream cache used to
train the Stream(Ours) security checkpoint) in the paragen_llm
security/prompt-injection-robustness experiments (Table 3: TensorTrust,
Gandalf, Purple, RuLES, StruQ, NESSiE, IFEval).

• Format: JSONL, one object per line, fields instruction / input / output
  (standard Alpaca schema).
• Size: 48,538 examples.
• Field mapping used downstream: instruction -> system message (falls
  back to a default system prompt if empty), input -> user message (may be
  empty), output -> assistant target.
• Provenance: derived/regenerated from Alpaca-style instruction data;
  response text follows a fixed scaffold ("let me help you with this: ...
  Therefore, the complete response is as follows: ...").

License note

This is Alpaca-derived instruction data. Upstream Alpaca-style datasets are
typically released under CC-BY-NC-4.0 due to originating from an OpenAI-model
completion pipeline, which restricts commercial use. Treat this dataset under
the same restriction unless you can independently verify otherwise.

Related code

Preprocessing into the multi-stream training cache (2D position_ids +
cross-channel attention mask) is implemented in
train/custom_datasets/process_dataset_fastresponse_safe.py in the
paragen_llm repo.