--- dataset_info: - config_name: all_organism_annotation features: - name: entry dtype: string - name: target_relation dtype: string - name: instruction dtype: string - name: reasoning dtype: string - name: answer dtype: string splits: - name: train num_bytes: 2248542263.378806 num_examples: 1718675 - name: validation num_bytes: 86197336.33334553 num_examples: 65885 - name: test num_bytes: 23622662.287848327 num_examples: 18056 download_size: 434648391 dataset_size: 2358362262.0 - config_name: all_organism_generation features: - name: entry dtype: string - name: target_relation dtype: string - name: target_value dtype: string - name: instruction dtype: string - name: reasoning dtype: string - name: answer dtype: string splits: - name: train num_bytes: 6147186072.329801 num_examples: 2941547 - name: validation num_bytes: 217454147.0669514 num_examples: 104056 - name: test num_bytes: 62695489.603248335 num_examples: 30001 download_size: 1309929051 dataset_size: 6427335709.000001 - config_name: ppi_annotation features: - name: entry_a dtype: string - name: entry_b dtype: string - name: instruction dtype: string - name: reasoning dtype: string - name: answer dtype: string - name: label dtype: string splits: - name: train num_bytes: 282097207.1774354 num_examples: 101524 - name: validation num_bytes: 1856122.043994788 num_examples: 668 - name: test num_bytes: 6068518.778569786 num_examples: 2184 download_size: 77352111 dataset_size: 290021847.99999994 - config_name: ppi_generation features: - name: entry_a dtype: string - name: entry_b dtype: string - name: instruction dtype: string - name: reasoning dtype: string - name: answer dtype: string splits: - name: train num_bytes: 117502225.60757262 num_examples: 50762 - name: validation num_bytes: 773132.3303441404 num_examples: 334 - name: test num_bytes: 2527726.0620832373 num_examples: 1092 download_size: 36837122 dataset_size: 120803084.0 - config_name: scop_annotation features: - name: entry dtype: string - name: target_relation dtype: string - name: instruction dtype: string - name: reasoning dtype: string - name: answer dtype: string splits: - name: train num_bytes: 35778372.65808565 num_examples: 33024 - name: validation num_bytes: 3402975.669787035 num_examples: 3141 - name: test num_bytes: 2483164.672127311 num_examples: 2292 download_size: 8092078 dataset_size: 41664513.0 - config_name: scop_generation features: - name: entry dtype: string - name: target_relation dtype: string - name: target_value dtype: string - name: instruction dtype: string - name: reasoning dtype: string - name: answer dtype: string splits: - name: train num_bytes: 41281902.13993774 num_examples: 32141 - name: validation num_bytes: 3966227.367167411 num_examples: 3088 - name: test num_bytes: 2204030.4928948437 num_examples: 1716 download_size: 12338288 dataset_size: 47452160.0 - config_name: task_a_annotation features: - name: entry dtype: string - name: target_relation dtype: string - name: instruction dtype: string - name: reasoning dtype: string - name: answer dtype: string splits: - name: train num_bytes: 55844564.66212483 num_examples: 35642 - name: validation num_bytes: 4058061.3454605048 num_examples: 2590 - name: test num_bytes: 7022482.992414665 num_examples: 4482 download_size: 13900695 dataset_size: 66925109.0 - config_name: task_a_generation features: - name: entry dtype: string - name: target_relation dtype: string - name: target_value dtype: string - name: instruction dtype: string - name: reasoning dtype: string - name: answer dtype: string splits: - name: train num_bytes: 256133960.82203332 num_examples: 111732 - name: validation num_bytes: 19022299.85054624 num_examples: 8298 - name: test num_bytes: 29312864.327420436 num_examples: 12787 download_size: 62508605 dataset_size: 304469124.99999994 configs: - config_name: all_organism_annotation data_files: - split: train path: all_organism_annotation/train-* - split: validation path: all_organism_annotation/validation-* - split: test path: all_organism_annotation/test-* - config_name: all_organism_generation data_files: - split: train path: all_organism_generation/train-* - split: validation path: all_organism_generation/validation-* - split: test path: all_organism_generation/test-* - config_name: ppi_annotation data_files: - split: train path: ppi_annotation/train-* - split: validation path: ppi_annotation/validation-* - split: test path: ppi_annotation/test-* - config_name: ppi_generation data_files: - split: train path: ppi_generation/train-* - split: validation path: ppi_generation/validation-* - split: test path: ppi_generation/test-* - config_name: scop_annotation data_files: - split: train path: scop_annotation/train-* - split: validation path: scop_annotation/validation-* - split: test path: scop_annotation/test-* - config_name: scop_generation data_files: - split: train path: scop_generation/train-* - split: validation path: scop_generation/validation-* - split: test path: scop_generation/test-* - config_name: task_a_annotation data_files: - split: train path: task_a_annotation/train-* - split: validation path: task_a_annotation/validation-* - split: test path: task_a_annotation/test-* - config_name: task_a_generation data_files: - split: train path: task_a_generation/train-* - split: validation path: task_a_generation/validation-* - split: test path: task_a_generation/test-* --- # eshmun-thinking Thinking-aware instruction-tuning dataset for protein language models: chain-of-thought `...` reasoning traces, built **programmatically** from UniProt/SCOP knowledge graphs (not distilled from an external LLM or self-sampled), covering protein **annotation** (sequence → description) and **generation** (description → sequence) in both directions. Hypothesis under test: does inserting an explicit reasoning step grounded in real sequence features (motifs, domains, GO function/process/localization, EC numbers, SCOP fold/superfamily/ family, PPI partners) before the final answer improve instruction-tuned protein models over direct instruction tuning (InstructProtein/ProLLaMA-style)? ## Configs 8 configs, each with `train`/`validation`/`test` splits (sequence-identity-based via MMseqs2: validation <70% identity to train, test <30% identity to train): | config | task family | direction | examples | |---|---|---|---| | `task_a_annotation` | family/function/process/localization/catalytic activity (human) | sequence → property | 42,714 | | `task_a_generation` | same, generation direction | property → sequence | 132,817 | | `all_organism_annotation` | same task family, all reviewed SwissProt organisms | sequence → property | 1,802,616 | | `all_organism_generation` | same, generation direction | property → sequence | 3,075,604 | | `ppi_annotation` | protein-protein interaction (1:3 positive:negative) | pair → yes/no | 104,376 | | `ppi_generation` | protein-protein interaction | partner + context → sequence | 52,188 | | `scop_annotation` | SCOP fold/superfamily/family | sequence → property | 38,457 | | `scop_generation` | SCOP fold/superfamily/family | property → sequence | 36,945 | Each row has `instruction`, `reasoning` (the `...` block), `answer`, plus task-specific identifier columns (`entry`, `entry_a`/`entry_b`, `target_relation`, `target_value`, `label`). ## Protein sequence encoding Sequences are wrapped as `...` with a `Ƥ` character joining every residue (`"Ƥ".join(sequence)`), e.g. `MƤSƤLƤGƤLƤL...`. This matches the tokenization convention of the InstructProtein-based pilot this dataset currently trains (`khairi/Eshmun-Thinking-Pilot`), whose tokenizer has no protein-specific vocabulary and would otherwise BPE-merge multiple residues into one token, losing residue-level granularity. Strip `Ƥ` to recover the raw sequence. ## Reasoning trace subjects **Annotation direction**: reasoning cites real, known facts about the specific queried protein (or, for SCOP, the other two structural levels of the same solved entry) — subject is literally `protein` (or `protein_a`/`protein_b` for PPI, since both are given inputs). **Generation direction**: the model is given only one stated property as input (e.g. `member_of = "Cytochrome P450 family"`), so reasoning cannot cite other instance-specific facts about a not-yet-generated sequence without asserting things the model can't know and inverting the real causal order (sequence determines function, not the reverse). Instead it cites **population-level statistics about the target category**, computed from train-split entries only, and the triple subject reflects that — a short (one/two-word) category label, not `protein`: | `target_relation` | subject | |---|---| | `catalyzes` | `catalytic_activity` | | `has_function` | `molecular_function` | | `involved_in` | `biological_process` | | `located_in` | `cellular_component` | | `member_of` | `family` | | `scop_fold` | `fold` | | `scop_superfamily` | `superfamily` | | `scop_family` | `scop_family` | Example (`task_a_generation`, target `catalyzes = 1.-.-.-`): ``` instruction: I want a protein sequence with the catalytic activity described by EC number 1.-.-.-. reasoning: (catalytic_activity, has_length, [>400] amino acids) (catalytic_activity, involved_in, glycolipid biosynthetic process) (catalytic_activity, located_in, mitochondrion) ... answer: MƤLƤPƤNƤTƤGƤRƤL... ```