open-athena/pdbthink-coordinate-tasks

PDBThink Coordinate Tasks

100,000 new coordinate-interpretation tasks across all 19 active PDBThink families, from 2,671 experimental PDB entries in 1,867 source groups. Version 1.3.0; deterministic seed 2026100101.

The model receives sanitised, rotated, rounded protein coordinates and a question. It must answer without tools. This release contains no sequence-to-structure prediction tasks and no retired MECH tasks. It is intended for additional evaluation, RL with deterministic rewards, and generating oracle-checked SFT demonstrations.

Revision history

• v1.3.0: retains all 100,000 v1.2.0 task identities, displayed coordinates, gold answers and grouped splits. Prompt v5 states the clash exclusions and ranking rule for G04. Scorer 1.1.0 compares the decimal representations of parsed numeric values exactly at the inclusive tolerance boundary; it adds no tolerance slack. Both coordinate triples and distances use this rule.
• v1.2.0: expands the corrected v1.1.0 release from 10,000 to 100,000 tasks.
• v1.1.0: corrects the S03/S05/S09 answer-format examples with prompt v4.

The release comparison and native-verifier regression evidence are in audits/contract_revision/. The geometry definitions and gold labels are unchanged. Historical releases remain available by their version tags. The published GLM teacher traces were generated and scored against v1.2.0; their prompts, scores and retry histories have not been rewritten. In particular, they do not constitute new evaluations of the clarified G04 prompts. Use the verifier bundled with each task release.

Relationship to the benchmark

These are new tasks, not the published benchmark questions. Acquisition excluded every source configured in the frozen benchmark inventory, exact full-entity and observed-chain protein sequences, and RCSB 30% sequence clusters associated with those sources. Every protein partner in a candidate entry was audited, even when only one chain was shown. The frozen inventory, source checks and file hashes are in audits/ and manifest.json. All released source, sequence and cluster intersections with that inventory are zero. This does not establish absence from foundation-model pretraining.

Gold answers are recomputed from the coordinates actually shown, after transformation, rounding and permitted cropping, and checked again after parsing the displayed PDB text. Existing PDBThink operational definitions and family oracles are used unchanged. Scientific ambiguities and failed generation criteria are recorded in ledger.parquet. This is an automatically verified training/evaluation release, not a curator-reviewed benchmark.

Format

The archive columns follow Open Athena Task Trove: path, source, family, template_id, converter, mode, dockerfile_id, language, tags, has_solution, task_binary, and solution_binary. Each binary is a gzip-compressed tar archive with deterministic member order and timestamps.

task_binary contains instruction.md, task.toml, prompt.json, environment/Dockerfile, tests/test.sh, tests/verifier.toml, and a self-contained Python verifier with hidden gold. Oracle solutions are separate in solution_binary, never baked into the environment. The custom pdbthink-coordinate verifier is bundled; it is not a built-in tasktrove-verify mode. The verifier writes exact correctness to /logs/verifier/reward.txt and diagnostic partial credit to details.json. Contact sets must be completely correct for unit reward.

Additional columns include split, source_group, input_tokens, atom_count, answer_schema, prompt_sha256, semantic_key, task_sha256, source_entries, and provenance. Provenance and test/solution files are evaluator-only and must never be included in a model prompt. Token counts use cl100k_base as a reference; compute exact native token counts before selecting an evaluation cohort or assigning output budgets.

from datasets import load_dataset
tasks = load_dataset("open-athena/pdbthink-coordinate-tasks", split="train")
contacts = tasks.filter(lambda row: row["family"] in ["S06", "I01"])

Tool-free execution

Ordinary terminal agents do not satisfy this scientific protocol. The source snapshot includes pdbthink.taskgen.no_tools_agent:CoordinateNoToolsAgent, a Harbor adapter which sends only the system/user prompt, explicitly disables model tools and the OpenRouter web plugin, retains raw responses, audits tool events, and writes the returned answer to /app/answer.txt on the host's behalf. It never executes a model-generated command or tool call. Use an exact per-prompt budget manifest containing native input counts, context and output limits, tokenizer revision and endpoint-limit provenance. The adapter allocates the largest supported output budget that fits. Truncated responses remain flagged for budget audit.

For offline oracle validation only, unpack solution_binary beside the task and run Harbor's oracle agent. Do not unpack oracle solutions into an evaluated model's environment. For SFT generation, provide only prompt.json to the teacher; check its response using the same verifier. Oracle answer files are final answers, not teacher-written worked solutions.

Splits and composition

All entries connected by exact protein sequences or RCSB 30% clusters stay in one split. The target allocation is 90% train / 5% validation / 5% test at source-group level, so task counts need not have those exact proportions. Do not randomly split rows or rotate a training structure and call it an independent held-out example. Multiple tasks may share a source structure; no task is counted solely because of a different rigid rotation.

Split	Tasks
test	4435
train	91154
validation	4411

Family	Tasks
G01	9795
G02	9795
G03	9795
G04	232
I01	291
N01	9793
P01	2823
P02	2526
P03	9795
S01	9795
S02	534
S03	9795
S04	9795
S05	1862
S06	407
S07	506
S08	1190
S09	9794
T01	1477

The selector balances families subject to the available unambiguous, non-overlapping source pool. Family counts and categorical label distributions are recorded in the manifest.

Reproduction and licence

generator_source.tar.gz contains the Python package, versioned definitions and project metadata. reproduction_inputs.tar.gz contains the frozen source selections, exclusion inventory, sequence-cluster snapshots and acquisition audits; unpack it into the work directory. Acquisition is the network stage; generation, verification and packaging run offline. Source selections, raw-source hashes and exclusion snapshots are retained for audit. See the bundled pdbthink.taskgen CLI and manifest.json for build settings and code hashes.

Task-generation code is distributed under the repository's Apache 2.0 licence. Protein coordinates come from the public Protein Data Bank; original entries are identified in evaluator-only provenance. No assertion of experimental folding ground truth is made beyond the supplied coordinate-reasoning definitions.

Validation and Snowball context

All 100,000 packaged tasks passed coordinate-oracle recomputation and solution verification. The source suite passed 397 tests. The correction audit executed the actual packaged verifier in Docker: both reported numeric boundary answers now pass, nearby outside-tolerance controls fail, and the clarified clash task accepts its gold pair while rejecting the excluded SG–SG pair. See checks.json, validation.json, and the regression evidence.

The exact Snowball tokenizer/chat-template recount leaves 28,045 training tasks within the 32,768-token context with 8,192 tokens available for reasoning and output: the same task identities as v1.2.0. Across all splits the count is 30,418. The longer G04 wording reduces prompt-only fit from 48,760 to 48,759 tasks and 4K-headroom fit from 38,458 to 38,457. The full counts and per-task prompt hashes are in snowball_context.json and snowball_context.parquet. These are cohort filters, not generation caps.

The frozen source files match GitHub commit 0a49f3b; see code-provenance.json and SHA256SUMS. Earlier independent reviews remain under audits/independent_review/ with an explicit historical-evidence note.