v2: retrained on scaled+deduplicated benchmark (host 0.995/0.993, engineered 0.909/0.874, 5-class 0.69); add read-filter CLI

README.md

@@ -26,10 +26,10 @@ model-index:
       split: test
     metrics:
     - type: roc_auc
-      value: 0.982
+      value: 0.995
       name: AUROC (test)
     - type: roc_auc
-      value: 0.937
+      value: 0.993
       name: AUROC (novel taxa)
   - task:
       type: text-classification
@@ -40,10 +40,10 @@ model-index:
       split: test
     metrics:
     - type: roc_auc
-      value: 0.895
+      value: 0.909
       name: AUROC (test)
     - type: roc_auc
-      value: 0.806
+      value: 0.874
       name: AUROC (novel taxa)
   - task:
       type: text-classification
@@ -54,13 +54,13 @@ model-index:
       split: test
     metrics:
     - type: accuracy
-      value: 0.65
+      value: 0.692
       name: Accuracy (test, 5-class)
 ---
 
 # dna-origin-classifier
 
-A reference-free, alignment-free classifier that labels a DNA coding sequence by its source:
+A reference-free, alignment-free classifier that labels a DNA sequence by its source:
 human, other eukaryote, bacterial, viral, or engineered/synthetic. It uses no alignment and no
 sequence database. One fixed k-mer featurizer feeds three linear heads.
 
@@ -90,22 +90,34 @@ clf.host_score(seq)        # higher = more human/host-like
 clf.engineered_score(seq)  # higher = more likely engineered/synthetic
 ```
 
-Requires only `numpy` and `safetensors`.
+A read-filter CLI (`dna_filter.py`) wraps the host head for FASTQ/FASTA input in two modes:
+host depletion (pathogen enrichment) and human removal (privacy). Requires only `numpy` and
+`safetensors`.
 
 ## Evaluation
 
 Measured from the published weights on the test and novel-taxa splits of
-[dna-origin-benchmark](https://huggingface.co/datasets/phanerozoic/dna-origin-benchmark):
+[dna-origin-benchmark](https://huggingface.co/datasets/phanerozoic/dna-origin-benchmark)
+(24,950 fragments, 382 organisms, cluster-level splits):
 
 | task | head | test | novel taxa |
 |---|---|---|---|
-| human vs non-host | `host` | 0.982 | 0.937 |
-| engineered vs natural | `engineered` | 0.895 | 0.806 |
+| human vs non-host | `host` | 0.995 | 0.993 |
+| engineered vs natural | `engineered` | 0.909 | 0.874 |
 
-Five-class origin accuracy on the held-out test split: 0.65 (random baseline 0.20). The
+Five-class origin accuracy on the held-out test split: 0.692 (random baseline 0.20). The
 human/eukaryote boundary is the hardest case and accounts for most of the five-class error; the
 binary `host` head is the right tool when the question is human versus non-host.
 
+## Comparison with database tools
+
+Run on the same splits, Kraken2 against the RefSeq Standard-8 database reaches host sensitivity
+0.972 and specificity 1.000 on databased taxa, so on organisms present in a database it is at
+least as strong. The difference is on sequence absent from every database (the benchmark's
+`undatabased` split): Kraken2 classifies 0% of it and BLAST 6.6%, while this classifier calls
+100% of it (97% as non-human), because it needs no reference. It is also reference-free and about
+50,000x smaller than an 8 GB database index.
+
 ## Calibration
 
 The heads output linear margins, not calibrated probabilities. Use the argmax of `logits` for
@@ -114,15 +126,16 @@ read the raw values as probabilities.
 
 ## Training data
 
-All heads are fit on the train split of `dna-origin-benchmark`. Sequences are RefSeq coding
-sequences and NCBI synthetic-construct records.
+All heads are fit on the train split of `dna-origin-benchmark`: RefSeq coding sequences across
+~370 taxa, NCBI synthetic-construct and vector records, synonymous recodings, and environmental
+metagenome fragments.
 
 ## References
 
 - **Benchmark, splits, and baselines:** [phanerozoic/dna-origin-benchmark](https://huggingface.co/datasets/phanerozoic/dna-origin-benchmark).
 - **Comparison model:** [HuggingFaceBio/Carbon-8B](https://huggingface.co/HuggingFaceBio/Carbon-8B), an 8B-parameter genomic language model evaluated zero-shot on the same splits. This classifier shares no weights or outputs with it and is not a derivative.
 - **Method lineage:** k-mer naive-Bayes sequence classification, as in the RDP Classifier (Wang et al., 2007), and k-mer "genomic signatures" of composition (Karlin & Burge, 1995).
-- **Reference-based tools compared on the benchmark:** Kraken2 (taxonomic classification against a sequence database) and Synsor (alignment-free engineered-DNA detection).
+- **Database tools compared:** Kraken2 (taxonomic classification against a sequence database) and Synsor (alignment-free engineered-DNA detection).
 
 ## License