MIMIC 1.0

README.md

@@ -1,3 +1,82 @@
+<!--
+TODO before making the repo public:
+  - Set the weights license below (the CODE is Apache-2.0; the weights license is
+    Polymathic's call). Add `license:` (+ `license_name`/`license_link` if "other")
+    to the YAML front-matter.
+  - Fill in the Training data and Citation sections.
+  - Update the install instructions once the `mimic` package location is final.
+-->
 ---
-license: mit
+tags:
+  - biology
+  - genomics
+  - dna
+  - rna
+  - protein
+  - multimodal
+  - foundation-model
 ---
+
+# MIMIC 1.0
+
+MIMIC is a multimodal encoder–decoder foundation model of the central dogma, trained
+jointly over **DNA, RNA, and protein** together with a range of structural and
+functional tracks. A single model embeds any subset of modalities into a shared
+representation space and generates any modality conditioned on the others.
+
+- **Architecture:** 20-layer encoder × 1536-d, 12-layer decoder × 1536-d, rotary
+  position embeddings, mixed (uni-/bi-directional) attention, 5 register tokens.
+- **Parameters:** ~1.25B.
+- **Weights:** `bfloat16`, `safetensors`.
+- **Modalities (26):** nucleotide/amino-acid sequence, codons, splice junctions/regions,
+  coding annotation, conservation (phyloP human/mouse), protein structure tokens, DSSP,
+  SASA, MaSIF surface features, protein abundance, RASP2 reactivity, and free-text
+  functional/context channels.
+
+## Usage
+
+```python
+from mimic import load_pretrained
+
+# Downloads config.json + model.safetensors pinned to git tag v1.0.
+model = load_pretrained(version="1.0")
+
+# --- Embed ---
+model.input([{"rna_seq": "ACGUACGUACGUACGU"}])
+reps = model.embed()                 # per-sample encoder representations
+#   reps[0]["full"]      -> (num_tokens, 1536) tensor
+#   reps[0]["dna/rna"]   -> just the RNA tokens
+
+# --- Generate ---
+# Mask positions you want the model to fill (or request a co-grouped modality),
+# then generate. Default strategy is an Ensemble soft-vote (deterministic at low temp).
+model.input([{"rna_seq": "ACGU____"}])
+out = model.generate("rna_seq", strategy="one_shot")
+print(out["rna_seq"]["preds"])       # detokenized prediction
+```
+
+`load_pretrained` fetches only `config.json` + `model.safetensors`; tokenizers ship
+inside the `mimic` package.
+
+> **Install:** the `mimic` package is distributed separately.
+> `TODO: pip install mimic` / link to the package repository.
+
+## Files
+
+| File | Description |
+|------|-------------|
+| `config.json` | Architecture + modality configuration consumed by `load_pretrained`. |
+| `model.safetensors` | Model weights (bf16). |
+
+## Training data
+
+`TODO: summarize the pretraining corpus (sources, modalities, scale).`
+
+## Citation
+
+`TODO: add citation / BibTeX.`
+
+## License
+
+`TODO: confirm the weights license.` The accompanying source code is released under
+the Apache License 2.0.

config.json

@@ -0,0 +1,89 @@
+{
+  "encoder": "xt_encoder_20L_1536D",
+  "decoder": "xt_decoder_12L_1536D",
+  "encoder_dim": null,
+  "decoder_dim": null,
+  "encoder_depth": null,
+  "decoder_depth": null,
+  "encoder_position_code": "rotary",
+  "decoder_position_code": "rotary",
+  "rotary_emb_ratio": 0.75,
+  "encoder_use_alibi": false,
+  "decoder_use_alibi": false,
+  "attn_flash": true,
+  "mixed_attention": true,
+  "unidir_attention_ratio": 0.5,
+  "num_register_tokens": 5,
+  "sum_modality_groups": true,
+  "exclude_absent_tokens": true,
+  "decoder_causal_mask": false,
+  "decoder_sep_mask": true,
+  "share_model_embeddings": true,
+  "drop_enc_rate_min": 0.0,
+  "drop_enc_rate_max": 0.1,
+  "class_balance_max": null,
+  "freeze_llm_emb": true,
+  "num_input_tokens": 10000,
+  "num_target_tokens": 1000,
+  "is_target_autoregr": false,
+  "dtype": "bfloat16",
+  "in_domains": [
+    "tok_aa_seq",
+    "tok_atac",
+    "tok_cage",
+    "tok_cds_junctions",
+    "tok_context",
+    "tok_corpus",
+    "tok_dssp",
+    "tok_feature_type",
+    "tok_funcprot_caption",
+    "tok_gene_family_txt",
+    "tok_is_coding",
+    "tok_masif_charge",
+    "tok_masif_hbond",
+    "tok_masif_hydrophobicity",
+    "tok_masif_n_vertices",
+    "tok_masif_si_index",
+    "tok_phylop_human",
+    "tok_phylop_mouse",
+    "tok_prot_abund",
+    "tok_prot_struct",
+    "tok_rasp2",
+    "tok_rna_codons",
+    "tok_rna_seq",
+    "tok_sasa",
+    "tok_splice_jctns_5cls",
+    "tok_splice_regions"
+  ],
+  "out_domains": [
+    "tok_aa_seq",
+    "tok_atac",
+    "tok_cage",
+    "tok_cds_junctions",
+    "tok_context",
+    "tok_corpus",
+    "tok_dssp",
+    "tok_feature_type",
+    "tok_funcprot_caption",
+    "tok_gene_family_txt",
+    "tok_is_coding",
+    "tok_masif_charge",
+    "tok_masif_hbond",
+    "tok_masif_hydrophobicity",
+    "tok_masif_n_vertices",
+    "tok_masif_si_index",
+    "tok_phylop_human",
+    "tok_phylop_mouse",
+    "tok_prot_abund",
+    "tok_prot_struct",
+    "tok_rasp2",
+    "tok_rna_codons",
+    "tok_rna_seq",
+    "tok_sasa",
+    "tok_splice_jctns_5cls",
+    "tok_splice_regions"
+  ],
+  "init_text_from_biobert": false,
+  "mimic_version": "1.0",
+  "source_checkpoint": "5.7_main/E20-D12-H1536-pt7-synced/checkpoint-final.pth"
+}

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:75d09f50b5343fe243c1c87123adf56a08cff8d55ff80749fd50bc6f963f6dc9
+size 2824113784