CoolFace
Datasetpublic

Hkang/endless-terminals-qwen3-8b-embeddings-32k

Endless Terminals — Qwen3-8B state/action embeddings All 44,128 released trajectories from 2,492 tasks, yielding 156,941 state/action pairs. Includes o3 and Llama-3.2-3B-Instruct attempts, both successful and failed. No new rollouts were generated. The upstream dataset card declares MIT; original licenses remain applicable. The upstream generator code is Apache-2.0. Normalized untruncated source with exact original file provenance: Hkang/endless-terminals-normalized-trajectories… See the full description on the dataset page: https://huggingface.co/datasets/Hkang/endless-terminals-qwen3-8b-embeddings-32k.

sourceHugging Facemitupdated 7d agoView on Hugging Face
0likes46downloads
Dataset Card

Endless Terminals — Qwen3-8B state/action embeddings

All 44,128 released trajectories from 2,492 tasks, yielding 156,941 state/action pairs. Includes o3 and Llama-3.2-3B-Instruct attempts, both successful and failed. No new rollouts were generated. The upstream dataset card declares MIT; original licenses remain applicable. The upstream generator code is Apache-2.0.

Normalized untruncated source with exact original file provenance: Hkang/endless-terminals-normalized-trajectories at 7f0231e242e00c94f207323d9f55c80ae8a37d83. Encoder: Qwen/Qwen3-8B at b968826d9c46dd6066d109eabc6255188de91218.

Each state is the original conversation preceding its action, with text and role order preserved. The source records terminal observations as user messages. Each nonempty assistant message is a target, including completion signals and malformed commands. Recorded reasoning text is preserved. Outcomes are numeric 1/0 at trajectory level, not individual-step correctness labels. Verifier results remain provenance only and are never included in model inputs.

Configured context: 32,768, without YaRN. States retain the last 32,767 tokens; actions retain the first 32,767. This can remove early instructions from long states. Truncated states: 38; truncated actions: 0. Original lengths and removed token counts are retained per pair.

Final-layer LAST-token pooling, L2 normalization, 4,096 dimensions, BF16 inference, FP16 saved vectors, no projection head. train/ and val/ each contain state_embeddings.pt, action_embeddings.pt, and aligned metadata.json. Original task IDs are retained; a deterministic task-ID hash keeps every attempt and model for a task in one split. This is exact task separation, not semantic deduplication. Trajectory IDs identify source file and attempt; source message indices remain in metadata. This covers the publicly released task snapshot; additional unreleased tasks described in the paper are excluded.

manifest.json pins source and encoder revisions. validation.json reports exact coverage, disjoint tasks, checkpoint/output checksums, finite unit-normalized vectors and clipping counts. pipeline/ contains reproduction code.