kader-xai/priya-sft

Priya SFT + RAG dataset

Synthetic training and retrieval data for the Priya persona — a fictional Senior Customer Success Manager at a fictional B2B SaaS company. Used to train kader-xai/priya-qwen2.5-7b-lora and kader-xai/priya-qwen2.5-7b-gguf.

Fully synthetic. No real person, customer, or company. Generated as the seed corpus for Project Recall, an experiment in employee-continuity AI.

• 📝 Blog post: Employee Recall — Capturing a Departing Employee's Writing Style and Memory
• Source code: github.com/kader-xai/EmployeeRecall
• How this was built: train-own-persona.md

---

Files

File	Lines	Purpose
sft_train_priya.jsonl	1,223	SFT training pairs (chat format)
sft_eval_priya.jsonl	64	SFT held-out eval pairs
rag_docs_priya.jsonl	17,663	Chunked corpus for RAG retrieval
system_prompt_priya.txt	—	Voice-fingerprint system prompt used at inference

SFT format

Each line is a chat-format conversation: a system prompt (Priya's voice fingerprint), a user turn (the prior email thread), and an assistant turn (Priya's actual reply).

{
  "messages": [
    {"role": "system", "content": "You are an AI assistant that drafts email replies in the voice of Priya Sharma..."},
    {"role": "user",   "content": "From: customer@acme.com\nSubject: Re: Q3 renewal\n\nHi Priya — circling back on the pricing..."},
    {"role": "assistant", "content": "Hi Mike,\n\nThanks for the patience. To flag a couple of things..."}
  ]
}

Pairs are reconstructed from the synthetic corpus by grouping emails into threads, sorting by thread_position + date, and turning every Priya-authored reply (with prior context) into one (thread → reply) pair. 95/5 train/eval split.

RAG format

Each line is a chunk of the corpus with provenance and metadata, ready to embed and load into FAISS / Qdrant / etc.

{
  "doc_id": "...",
  "doc_type": "email | meeting_notes | transcript | rfc | adr | postmortem",
  "chunk_id": "...",
  "text": "...",
  "metadata": {
    "date": "...",
    "from": "...",
    "subject": "...",
    "boost": 1.0
  }
}

Chunking is type-aware:

• Email: one chunk per email (split by paragraph if > 1500 chars). Preamble carries from/to/subject/date.
• Structured meeting notes: one chunk per section (discussion / decisions / actions). Decisions and actions get a higher retrieval boost.
• Transcripts: window of 8 segments per chunk with timestamps + speakers.
• RFC / ADR / design doc: split on markdown headings.

Recommended embedder: BAAI/bge-base-en-v1.5 (768-dim, normalized) → IndexFlatIP for ≤ 1M chunks.

How to use

from datasets import load_dataset

sft = load_dataset("kader-xai/priya-sft", "sft")
rag = load_dataset("kader-xai/priya-sft", "rag", split="docs")

print(sft["train"][0]["messages"])
print(rag[0]["text"])

End-to-end training recipe in the persona-training article.

Provenance & ethics

• 100% synthetic. Generated by scripts/generate_bulk_corpus.py and v2 in the source repo. Deterministic — same seed produces the same corpus.
• The persona is fictional; no real mailbox was used. If you adapt this pipeline for a real person, you need consent, scope agreement, and a retention policy (see the methodology doc in the source repo).

License

CC-BY-4.0. Use it freely for research, models, or examples — attribution appreciated.