DEBRIS tagger debris_v2
The production tagger behind debris.mdeller.com: a single LoRA
fine-tune of Qwen/Qwen2.5-3B-Instruct, fused into the base weights (bf16), that turns a
PubMed title, journal, MeSH list and abstract into the DEBRIS tag JSON: relevance,
polymers, size class, organ systems, tissues, disease categories, pathways, study type,
evidence rungs, PET specificity and a short summary. It is trained on the task and the
schema, not on facts, and the weekly pipeline never retrains it. Tags record what a paper
reports; nothing in the schema asserts human causation.
Training
| Base model | Qwen/Qwen2.5-3B-Instruct |
| Method | LoRA (rank 8, scale 20) on 32 layers, batch size 4, 1,600 iterations, learning rate 1e-5, loss on the answer only, seed 42 |
| Data | 1,438 hand-checked abstracts (the canonical training set), recipe v2: model types, pathways and polymers as revised in the lab (D50), tissues added |
| Prompt | the system prompt from pipeline/lab_v2.py (system_prompt with the v2 switches); the model expects it byte for byte |
Test split (162 papers, frozen, scored once)
| Field | Score |
|---|---|
| Relevant (F1) | 0.971 |
| Polymers (F1) | 0.942 |
| Organ systems (F1) | 0.795 |
| Pathways (F1) | 0.798 |
| Evidence rung, exact | 0.870 |
| Evidence rung, within one | 0.907 |
| Study type (accuracy) | 0.932 |
| Tissues (F1) | 0.807 |
| Valid JSON | 1.000 |
Mean of the eight floor metrics: 0.902.
Built by Marc C. Deller, D.Phil., Elora Therapeutics.
- Downloads last month
- 25
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐ Ask for provider support