54 lines
1.6 KiB
Markdown
54 lines
1.6 KiB
Markdown
---
|
|
base_model: openbmb/MiniCPM5-1B
|
|
library_name: transformers
|
|
license: apache-2.0
|
|
language:
|
|
- en
|
|
pipeline_tag: text-generation
|
|
tags:
|
|
- minicpm
|
|
- openbmb
|
|
- qlora
|
|
- llama.cpp
|
|
- structured-generation
|
|
- build-small-hackathon
|
|
datasets:
|
|
- build-small-hackathon/compliment-forest-sft
|
|
---
|
|
|
|
# Compliment Forest MiniCPM5-1B
|
|
|
|
This is the merged text fine-tune for
|
|
[The Compliment Forest](https://huggingface.co/spaces/build-small-hackathon/compliment-forest).
|
|
It turns a name and situation into schema-valid, situation-grounded forest JSON for the app's
|
|
author pass. The same model also runs a bounded critic pass that prunes generic or redundant
|
|
clearings.
|
|
|
|
## Training
|
|
|
|
- Base: `openbmb/MiniCPM5-1B` (Llama architecture, about 1.08B parameters)
|
|
- Data: `build-small-hackathon/compliment-forest-sft`
|
|
- Method: 4-bit NF4 QLoRA on Modal
|
|
- LoRA: rank 16, alpha 32, dropout 0.05
|
|
- Targets: attention and MLP projections
|
|
- Sequence length: 2,048
|
|
- Epochs: 2
|
|
- Learning rate: 2e-4 with cosine decay
|
|
- Runtime thinking mode: disabled for deterministic JSON generation
|
|
|
|
The dataset was filtered for JSON validity, concrete situation grounding, non-toxic positivity,
|
|
and short first-person spells. This model is for whimsical encouragement; it is not a therapist
|
|
or a substitute for professional support.
|
|
|
|
## Inference
|
|
|
|
Use the base model's chat template with `enable_thinking=False`. The app enforces the output with
|
|
Pydantic and retries malformed generations at most twice.
|
|
|
|
The repository also includes a `Q4_K_M` GGUF build for local `llama.cpp` inference.
|
|
|
|
## License
|
|
|
|
Apache-2.0, following the base model and project code. Dataset source licenses are documented on
|
|
the dataset card.
|