gwagon63/innerbloom-reader-1.5b-GGUF

Innerbloom reader 1.5B

A small model that reads one journal entry and returns its main emotion, how strong it is, a one-line summary, tags, themes and the people named. It runs locally through Ollama in about 2.6 seconds per entry on an 8 GB M1 Mac, in 1.9 GB.

It is the entry reader for Innerbloom, a local-first AI journal. Innerbloom uses it automatically when Ollama has it, and keeps a larger model for chat.

Install

ollama pull hf.co/gwagon63/innerbloom-reader-1.5b-GGUF:Q8_0
ollama cp hf.co/gwagon63/innerbloom-reader-1.5b-GGUF:Q8_0 innerbloom-reader:1.5b-q8-edited

The second line gives it the name Innerbloom looks for. Or, from an Innerbloom checkout: python scripts/get_reader.py.

Results

On Innerbloom's locked test split: 400 fictional entries, scored once, never used for any choice. All three models ran through Ollama on the same Mac, with the same prompt, temperature 0.3 and seed 7.

	qwen2.5:7b	This model before the edit	This model
Emotion macro-F1 (16 classes)	0.652	0.702	0.746
Emotion accuracy, defensible alternatives	0.795	0.825	0.853
Intensity error, 1-10 (MAE)	1.155	0.777	0.795
Failed readings	0	0	0
People F1	0.823	0.916	0.923
Invented names	9	3	2
Median seconds per entry	13.3	3.4	2.6
Size	4.7 GB	1.6 GB	1.9 GB

On the 402-entry validation split the same model scored 0.752, so the score held on unseen entries. A second person labelled 100 validation entries blind: Cohen's kappa 0.84 with the original labels, and 99% agreement allowing defensible alternatives.

How it was made

1. Fine-tuning. QLoRA (rank 16, alpha 32, learning rate 2e-4, early stopping on validation loss) of Qwen2.5-1.5B-Instruct on 1,199 fictional entries. Targets: the labelled emotion, intensity and names, plus a summary, tags and themes written by qwen2.5:7b. About 6.5 Colab A100 compute units.
2. Quantisation. Merged, converted to GGUF with llama.cpp and quantised to Q8_0. Q4_K_M lost too much (0.672 on validation).
3. Weight edit. The model over-picked some emotions and avoided others. A per-emotion bias was fitted on training entries only, by coordinate ascent on macro-F1 checked with 5-fold cross-validation, and written into the model: Qwen2.5-1.5B shares one table for reading and writing tokens, so the edit adds a separate output table and moves the rows for 11 emotions' first tokens along the mean final hidden state. Reading is untouched. Validation macro-F1 went from 0.711 to 0.752, and test from 0.702 to 0.746.

Code and reports: ml/train_emotion.py, ml/weight_editing/, reports/final-test.md in the Innerbloom repository.

Prompt

The model was trained and scored on Innerbloom's own prompt, build_prompt in innerbloom_core/analysis.py, which asks for compact JSON:

{"emotion": "...", "intensity": 1, "summary": "...", "tags": ["..."], "themes": ["..."], "people": ["..."]}

The 16 emotions are happy, sad, anxious, calm, grateful, frustrated, excited, reflective, hopeful, tired, angry, content, lonely, proud, overwhelmed and neutral. Other prompts or tasks are untested.

Limits

• Fictional, English entries only. Every training and test entry was invented; none came from a real journal. Real writing may differ.
• Sarcasm is the weak spot: about 4 of 9 sarcastic test entries read correctly, the same as the 7B.
• Not a safety or medical tool. It does not detect crisis risk. Innerbloom checks for that separately, and nothing here should be used to judge anyone's mental health.
• One writer wrote the entries and labels; one second reader checked a sample.

Licence

Apache 2.0, the same as the base model Qwen2.5-1.5B-Instruct. The training data was written for Innerbloom; no public dataset was used in this model.