YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
DEBRIS tagger debris_v1
A single LoRA fine-tune of Qwen/Qwen2.5-3B-Instruct that turns a PubMed title, journal,
MeSH list and abstract into the DEBRIS tag JSON (spec 5.4). It is trained on the task
and the schema, not on facts: the weekly pipeline never retrains it.
Training
| Base model | Qwen/Qwen2.5-3B-Instruct |
| Method | LoRA, 32 layers, batch size 4, 1600 iterations, learning rate 1e-5, seed 42 |
| Data | debris-train:v1 (1277 train, 161 valid, 162 test), stratified on relevance and model type |
| Vocabulary | config/vocab.toml schema version 1.0 |
| Labels | build session (no ANTHROPIC_API_KEY) |
| wandb | offline |
| Fused model | data/adapters/debris_v1_fused (de-quantised safetensors, Hugging Face layout) |
| Quantised | data/adapters/debris_v1_q4 |
Test-split metrics
| Field | Value | Acceptance floor |
|---|---|---|
| n | 162 | |
| relevant_f1 | 0.9711 | 0.95 |
| polymers_f1 | 0.8701 | 0.85 |
| organ_systems_f1 | 0.8043 | 0.85 |
| pathways_f1 | 0.6332 | 0.75 |
| disease_categories_f1 | 0.908 | |
| rungs_exact | 0.8457 | 0.7 |
| rungs_within_one | 0.8642 | 0.9 |
| model_type_accuracy | 0.8523 | 0.9 |
| model_type_n | 88 | |
| valid_json | 1.0 | 0.98 |
| seconds_per_paper | 1.81 | |
| model | /Users/dellboy/Documents/Vibe_Coding/debris/data/adapters/debris_v1_fused |
Quantised (q4) comparison
| Field | Value |
|---|---|
| n | 162 |
| relevant_f1 | 0.948 |
| polymers_f1 | 0.8111 |
| organ_systems_f1 | 0.6704 |
| pathways_f1 | 0.5363 |
| disease_categories_f1 | 0.8607 |
| rungs_exact | 0.7346 |
| rungs_within_one | 0.7469 |
| model_type_accuracy | 0.7841 |
| model_type_n | 88 |
| valid_json | 0.9877 |
| seconds_per_paper | 1.45 |
| model | /Users/dellboy/Documents/Vibe_Coding/debris/data/adapters/debris_v1_q4 |
Acceptance
Below floor: organ_systems_f1 0.804 < 0.85; pathways_f1 0.633 < 0.75; rungs_within_one 0.864 < 0.9; model_type_accuracy 0.852 < 0.9
Intended use and limits
DEBRIS tags literature, it does not adjudicate science. A tag records what a paper reports: a polymer detected, an association observed in a cohort, a mechanism shown in vitro. Nothing in the schema asserts human causation, and the evidence rungs are deliberately ordered so that a reader can see how far from a human outcome a given claim sits. Reviews carry no rungs. Papers the tagger is unsure about are routed to the review queue and shown as "needs review" rather than silently published.
- Downloads last month
- 28