Hkang/adp-v1-qwen3-8b-embeddings-32k
ADP Dataset V1 — Qwen3-8B state/action embeddings All 1,270,306 trajectories and 2,592,191 typed ADP actions in the original V1 corpus (13 datasets / 18 subdirectories). No sampling or duplication of alternate SFT formats; later corpus additions are outside this V1 snapshot. Original source-specific licenses apply; no additional rights to the underlying source data are granted. Normalized, untruncated source and original file provenance: Hkang/adp-v1-normalized-trajectories at… See the full description on the dataset page: https://huggingface.co/datasets/Hkang/adp-v1-qwen3-8b-embeddings-32k.
ADP Dataset V1 — Qwen3-8B state/action embeddings
All 1,270,306 trajectories and 2,592,191 typed ADP actions in the original V1 corpus (13 datasets / 18 subdirectories). No sampling or duplication of alternate SFT formats; later corpus additions are outside this V1 snapshot. Original source-specific licenses apply; no additional rights to the underlying source data are granted.
Normalized, untruncated source and original file provenance: Hkang/adp-v1-normalized-trajectories at 5b1955855e1bfa39c23a3cc7c462ede06f8460a0. Encoder: Qwen/Qwen3-8B at b968826d9c46dd6066d109eabc6255188de91218.
Each state contains preceding ADP events, encoded as canonical JSON in chronological Qwen chat messages. Each action is one typed API/code/message action, including all its text fields and recorded description/reasoning. No action appears in its own state. Agent-authored observations remain history, not additional targets. The Qwen template is told that reasoning is already inside the serialized action; literal thinking tags do not strip source text. Image content is explicitly omitted for the text-only encoder; all HTML/accessibility-tree text is retained before clipping.
Context and vectors
Configured context: 32,768; states retain the last 32,767 tokens and actions the first 32,767. This may remove early instructions from long states. Truncated states: 133,770; truncated actions:
- Original lengths and removed-token counts are saved for every pair.
Final-layer LAST-token pooling, L2 normalization, 4,096 dimensions, BF16 inference, FP16 saved vectors. No learned projection head. No per-action correctness labels are inferred; rewards remain null.
Files
train/ and val/ contain state_embeddings.pt, action_embeddings.pt, and metadata.json. Tensor row order matches metadata samples. Trajectory IDs include source subset and row ordinal, while original source IDs and event indices remain in metadata. Task groups hash all observations preceding the first typed action; identical initial context stays in the same deterministic split. This is a conservative textual grouping, not a claim of semantic task deduplication.
manifest.json pins the source, encoder and counts. validation.json verifies complete pair coverage, train/validation task disjointness, chunk checksums, 4,096-dimensional finite/unit-normalized vectors, and exported-file checksums. pipeline/ contains reproduction code.