README.md

metadata

license: odc-by
language:
  - en
tags:
  - embeddings
  - swe-chat
size_categories:
  - 1K<n<10K

SWE-Chat Coding-Agent Session Embeddings

Embeddings of SALT-NLP/SWE-chat coding-agent session transcripts, produced with amkdg/Qwen3-Embedding-8B-NVFP4 — 4096-d, L2-normalized float16 (cosine = dot product).

• 5,830 conversations → 8,663 vectors
• emb.npy — float16 [8663, 4096]
• meta.parquet — one row per vector, aligned with emb.npy: id, uuid, tag, chunk, n_chunks, count, source_ref
• manifest.json — counts and provenance

Usage

import numpy as np, pyarrow.parquet as pq

emb  = np.load("emb.npy", mmap_mode="r")           # [8663, 4096] float16
meta = pq.read_table("meta.parquet").to_pandas()   # one row per vector, aligned with emb

# A conversation = consecutive rows sharing one `uuid` (`chunk == 0` marks its start);
# conversations longer than 8192 tokens span several chunk-rows.
starts = meta.index[meta.chunk == 0]               # first row of each conversation

Source mapping

Each row carries source_ref, the locator back into SALT-NLP/SWE-chat — source_ref is the session_id.

ref = meta.iloc[0].source_ref      # -> the matching conversation in the source dataset

Notes

One embedding per agent session: a short grounding header (repo · domain · language · agent · files · success) followed by only the conversational user/assistant turns. Tool calls, progress events, file snapshots and commits are dropped as noise (only ~3.8% of raw turns are conversational). tag is the repo domain (application / devtools / library). The ~26 sessions without dialogue fall back to their prompt summary.