Label	Meaning	Risk signal in Gomi
`satisfaction`	Positive / constructive tone	No
`neutral`	Neutral tone	No
`caution`	Hedging, workaround, “revisit later”	Yes
`frustration`	Negative / stressed tone	Yes

At runtime, Gomi aggregates per-file sentiment score as
(frustration + caution) / non–low-info commits.

Training

Item	Detail
Base model	distilbert-base-uncased
Training data	GitRatBCSAD/gomi-datasets → `openreview/openreview_labeled_2k.csv`
Script	`scripts/train_sentiment.py` in the Gomi ML repo

Prototype — Early checkpoint. Pin a revision tag (e.g. v1.0.0) for reproducible runs.

Files in this repo

File	Description
`config.json`	Model config (4 labels)
`model.safetensors`	Fine-tuned weights
`tokenizer.json`	Tokenizer
`tokenizer_config.json`	Tokenizer config

Inference needs the files above. The checkpoints/ folder may be present from training uploads; it is not required at runtime.

Download

Hugging Face CLI

bash

pip install huggingface_hub
huggingface-cli login   # if private

hf download GitRatBCSAD/gomi-sentiment \
  --local-dir ./distilbert_sentiment \
  --revision v1.0.0

Python

python

from transformers import AutoModelForSequenceClassification, AutoTokenizer

repo = "GitRatBCSAD/gomi-sentiment"
revision = "v1.0.0"  # or None for main

tokenizer = AutoTokenizer.from_pretrained(repo, revision=revision)
model = AutoModelForSequenceClassification.from_pretrained(repo, revision=revision)

Git + LFS

bash

git lfs install
git clone https://huggingface.co/GitRatBCSAD/gomi-sentiment
cd gomi-sentiment && git lfs pull

Use with Gomi

In the Gomi ML project .env:

bash

GOMI_SENTIMENT_MODEL_REPO=GitRatBCSAD/gomi-sentiment
GOMI_SENTIMENT_MODEL_REVISION=v1.0.0
HF_TOKEN=your_token_if_private

Or download into scripts/datasets/distilbert_sentiment/ and run gomi.py without the env var.

Training data lives in gomi-datasets, not in this repo.

Related repos

Repo	Role
gomi-datasets	CSV / JIT training data
gomi-risk	Logistic regression + SHAP background

License

Apache 2.0. Respect licenses of distilbert-base-uncased and the OpenReview training data when redistributing.