Tessera-SFT-Gold / README.md
AIIT-Threshold's picture
card: add configs block so viewer shows all 1,031 rows (3 files, one train split)
828f720 verified
|
Raw
History Blame Contribute Delete
1.89 kB
metadata
license: apache-2.0
language:
  - en
tags:
  - sft
  - hand-written
  - alignment
  - tessera
pretty_name: Tessera SFT Gold
configs:
  - config_name: default
    data_files:
      - split: train
        path:
          - proto_buddy_gold_531.jsonl
          - selfturns_v1.jsonl
          - selfturns_v2.jsonl

Tessera SFT Gold

Human-written supervised fine-tuning data for Tessera 1B — every response typed by hand by the founder of AIIT-THRESHOLD.

This is the cleanest artifact in the release. There is no license ambiguity and no provenance question: a person wrote both sides. It is the alignment doctrine made concrete — if you want a model that says what it means, you write the examples yourself.

Contents

File Pairs What it is
proto_buddy_gold_531.jsonl 531 The core hand-typed gold set (ids 001–531) — identity, tone, honesty, refusal, everyday exchanges
selfturns_v1.jsonl 250 Hand-authored self-turn script, batch 1
selfturns_v2.jsonl 250 Hand-authored self-turn script, batch 2
Total 1,031 all human-written, both sides

Schema: {"messages": [{"role": "user", ...}, {"role": "buddy", ...}]}.

Provenance & policy

  • 100% human-authored. Not model-generated, not scraped, not transcribed from an AI conversation. Consistent with AIIT's training-data policy: never tokenize a conversation with a model; no synthetic text.
  • Some prompts were scaffolded with tooling assistance, but every assistant reply was written and approved by hand.

Intended use

A small, high-signal SFT set for giving a base model a stable voice and an honest, non-sycophantic posture. Pairs naturally with Tessera 1B, but usable with any base.

License

Apache-2.0. Use it, remix it, build on it.