[davanstrien]

A second embeddings recipe for large text runs. generate-embeddings.py stays the default.

What it does

• Starts a vLLM server in the Job on the pinned vllm/vllm-openai:v0.30.0 image, and sends it one text per request with saturate.
• The [tool.hf-jobs] header sets the image, a10g-small, a 2h timeout and HF_TOKEN, so the command needs no job flags.
• Writes one row per input row (id, the text column, embeddings, model, prompt_tokens, error). A run can be resumed by running the same command again.
• Cuts each text at the model's own maximum, which the server reports, unless --max-seq-len is set. A value above the model's maximum is lowered to it.
• Picks the prompt as generate-embeddings.py does: --prompt, then the model's registered document prompt, then the known-family table.
• Writes the dataset card with the standard reproduction section and the uv-script and hf-jobs tags.

generate-embeddings-vllm.py stays in the folder. Its docstring and the README row now name this recipe as its successor.

Measured

One A10G, 20,000 IMDb reviews of up to 300 words, embedding step only, one run each.

Model	generate-embeddings.py	this recipe
Qwen/Qwen3-Embedding-0.6B	96 rows/s	187 rows/s
BAAI/bge-m3	71 rows/s	276 rows/s

The server was ready 100 to 120 seconds after it started in those runs, so generate-embeddings.py is quicker for a few hundred rows.

Checks

100-row smoke Jobs, each with 100 of 100 rows, 0 error rows, unit-length vectors and 100 unique ids:

• Default model, header only, on a10g-small: cut at the model maximum (32768), longest text 1164 tokens.
• Default model on l4x1 (an earlier version with a fixed 512-token cut).
• intfloat/multilingual-e5-small with --max-seq-len 100000: lowered to 512, prefix passage: .
• BAAI/bge-m3 with --max-seq-len 128: longest text 128 tokens. This run used the committed file.

Records are in maintenance/verifications.jsonl. tools/build-catalog.py --verify-models and tools/recipe-status.py --check pass.

Review

A Codex review found four issues outside the default path. All are fixed in 16457ba:

• A resumed run that processed no rows wrote a card with dim None and 0 embedded. The dimension now comes from the server, and the card gives the rows of the last run and the rows written earlier.
• --id-column index and --id-column content were read as saturate id strategies. They are passed as a callable, as in lighton-ocr3-saturate.py.
• The card's reproduce command now carries every option that changes the result.
• The docstring says that a --retry-errors run keeps the old error record, and how to read one record per id.

Tested after the fix: the committed file run twice on the same output. The fresh run wrote 100 of 100 rows. The resumed run processed 0 rows, left 100 rows with 100 unique ids, and kept a valid card. The --id-column fix was not run on a Job.

Notes

• At the end of a run the vLLM log prints an EngineDeadError traceback. It appears after the last row is written, when the server is stopped.
• Both this PR and embeddings: header on generate-embeddings.py, and nine more declared models #189 append to maintenance/verifications.jsonl and regenerate maintenance/STATUS.md. The second one to merge needs a merge from main first.
• With this recipe, Qwen/Qwen3-Embedding-0.6B is declared on the Run on Jobs Space (30 to 31 models on main).

🤖 Generated with Claude Code

https://claude.ai/code/session_016isSPpNH4MzMcKDBpSwokh