Bio-Circuit

Public evaluation cohorts for three text-only, causal circuit interpretability tasks on Qwen3-4B. The release contains 48 tasks (16 per taskset), full attribution graphs, hidden evaluation cohorts, generation manifests, and SHA-256 checksums. There is no training split and no validated SFT trajectory collection in this release. Do not describe these 48 evaluation tasks as independent training data.

The dataset is downloadable without credentials. Live tool calls and reward computation require a CUDA GPU subject backend (the existing setup uses an A100-80GB). Precomputed graphs do not replace live causal interventions. This release is not eligible for a run that excludes GPU-dependent environments.

Tasksets

Config	Prime environment	Task	Test rows
visible	wazupsteve/circuit-supernodes	Infer a concept from evidence and identify causally necessary, specific features	16
rediscovery	wazupsteve/supernode-rediscovery	Rediscover features for a named semantic goal	16
restoration	wazupsteve/suppressed-features	Identify interventions that restore a corrupted subject's output	16

All tasksets use the circuit-null harness included in the environment package. The environment source and setup are at WazupSteve/bio-circuit, and the release package is distributed through Prime's Environments Hub.

Download and inspect

The Dataset Viewer and load_dataset expose lightweight task indexes. Each index row points to a full graph-bearing cohort in a compressed JSONL file; index rows alone are not runnable tasks or labeled model responses.

from datasets import load_dataset

tasks = load_dataset("amitprakash2005/bio-circuit", "visible", split="test")
print(tasks[0])

Download a full cohort file without authentication:

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    "amitprakash2005/bio-circuit",
    "data/visible_dataset/heldout-cohorts.jsonl.gz",
    repo_type="dataset",
    token=False,
    # For reproducibility, supply revision=<release commit SHA>.
)

The environment pins an immutable dataset commit and downloads its own taskset's file on first use. Compressed files total about 2.6 GiB; decompressed graphs and parsed task objects require considerably more memory. Read one cohort at a time with gzip.open(path, "rt"); do not load all graph-bearing files into memory.

Files and schema

• indexes/*.jsonl: task ID, family, goal, cohort counts, full data filename, zero-based row index, and SHA-256 of the compressed data file.
• data/visible_dataset/heldout-cohorts.jsonl.gz: visible evidence tasks.
• data/supernode_rediscovery/promptless-supernodes.jsonl.gz: rediscovery tasks.
• data/output_restoration/suppressed-features.jsonl.gz: restoration tasks.
• data/**/*.manifest.json: original generation provenance sidecars.
• inventory.json: verified file sizes, row counts, and compressed-file hashes.
• schema.json: the environment's JSON Schema for a full cohort row.

Each full row has family, optional goal, and three lists: example_prompts, heldout_prompts, and control_prompts. Each example has prompt, target_token, graph, and optional corruption. Graphs contain metadata, nodes, and links. Corruption entries are (layer, feature, position, factor). Stable intervention identities are layer_feature; graph nodes can also carry a token position. The environment's schema and parser define the exact types.

Original manifests are retained verbatim. In particular, the visible manifest's dataset field uses the historical name heldout-cohorts-v2.jsonl.gz; the published path is heldout-cohorts.jsonl.gz. Its checksum refers to the published bytes. Loopback backend URLs in manifests describe the generation setup, not a public endpoint.

Construction and validation

Subject model: Qwen/Qwen3-4B. Transcoders: mwhanna/qwen3-4b-transcoders. Graphs use Circuit Tracer with node threshold 0.8 and edge threshold 0.98. The manifests record sampling seeds, cohort construction, screening, generation policy, schema version, and aggregate graph-quality measurements.

Task families include two-hop reasoning, agreement, comparison, weekdays, path composition, modular arithmetic, rule composition, spatial composition, and Boolean composition. Every released row is checked against the environment's Pydantic schema, and row counts and hashes are checked against the original manifests before upload. This is structural/provenance validation, not a new GPU replay or a claim of successful agent performance.

Evaluation boundaries and limitations

Heldout prompts, control prompts, and corruption specifications are public to operators but must remain outside the evaluated policy's messages and tools. Prevent policy access to raw dataset files when running a meaningful evaluation; public availability is not a secrecy guarantee. The indexes also expose goals that some tasksets intentionally withhold from the policy.

The 16 tasks per family are a small research evaluation set, not a representative sample of all model behavior. Causal rewards depend on the subject model, transcoders, numerical stack, and backend configuration. Generation and model execution can be nondeterministic. Graph approximations and cohort selection introduce limitations described by the generation manifests and source code.

This release contains task inputs and grading data, not verified gold circuits or successful agent trajectories. SFT data generation needs policy rollouts, live GPU-backed grading, and a separately documented acceptance procedure.

License and attribution

The generated dataset is released under Creative Commons Attribution 4.0 International. Attribute Bio-Circuit and link to this dataset, noting modifications. Model weights, transcoders, and third-party software are not included and retain their respective licenses. Dataset licensing does not relicense those dependencies.