Datasets:
File size: 2,775 Bytes
5ea18fd 7729cff 5ea18fd 7729cff | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 | ---
license: apache-2.0
language:
- en
task_categories:
- text-generation
pretty_name: Compliment Forest SFT
size_categories:
- 1K<n<10K
tags:
- minicpm
- structured-generation
- synthetic-data
- positive-reframing
- build-small-hackathon
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: validation
path: data/validation-*
---
# Compliment Forest SFT
Compliment Forest SFT teaches a small language model to turn a `(name, situation)`
pair into a strict JSON forest of grounded encouragement. Each forest contains five
distinct creature-strength clearings, a situation-specific line, an agency-oriented
reflection, a short first-person spell, and a creature-only image prompt.
## Dataset Size
- Train: 1,350 records
- Validation: 150 records
- Seed: 42
- Language: English
Every row contains:
- `name`
- `situation`
- `source`
- `teacher_model`
- `messages`: MiniCPM-compatible system, user, and assistant chat messages
## Construction
The released v1 records come from a deterministic synthetic formatting layer over
50 hand-authored situations spanning work, study, moving, relationships, parenting,
creative work, wellbeing, grief-adjacent moments, money, social situations, and
leadership.
The data pipeline was informed and tested against:
- `SALT-NLP/positive_reframing` for growth, optimism, neutralizing, and impermanence
framing patterns.
- `Estwld/empathetic_dialogues_llm` for situation-grounded supportive register.
- `asuender/motivational-quotes` for short mantra analysis.
Cohere Command A was used during pilot synthetic generation. Its pilot outputs were
schema-validated and filtered, but no Cohere row passed the stricter final v1
groundedness gate. The published v1 rows therefore use `source=template_coverage`
and have no teacher-model dependency.
## Validation
The builder rejects:
- invalid or extra JSON fields;
- fewer than three or more than six proposed strengths;
- blank or overlong fields;
- spells that do not begin in the first person or exceed 12 words;
- crisis or acute-risk inputs;
- any clearing line that does not repeat at least one concrete situation term;
- duplicate `(name, situation)` identities.
The validation split is deterministic and disjoint.
## Intended Use
This dataset is intended for supervised fine-tuning and evaluation of
`openbmb/MiniCPM5-1B` for The Compliment Forest hackathon project. It may also be
useful for experiments in small-model JSON adherence and supportive rewriting.
## Limitations
The records are synthetic and structurally more regular than natural writing. They
are not clinical advice, therapy data, or a substitute for safety evaluation. The
runtime application adds a separate crisis guard and author-critic quality pass.
|