Reproducibility

A result cannot be checked later without its provenance, the record of how it was produced. This page covers section 01.07 of notebook 01: what to record with every result, and the code that records it.

What to record

Write these down with every result:

• Tag, for example nishad/apertus:8b.
• Digest, the fingerprint of the exact file and template.
• Quantization, for example Q4_K_M.
• Options, such as temperature, seed and num_ctx.
• Prompt, including the system message.
• Date of the run.
• Runtime version, the Ollama version.

Why

On the same runtime and hardware, the same digest at temperature 0 with a seed gives the same answer. Without the digest, you cannot say you used the same model: a tag can be pointed to a new file. A different quantization is a different model, with different answers. The runtime version belongs in the record too, so that you can trace a difference that appears later. Each item on the list removes one way that two runs could differ without anyone noticing.

This matters for metadata work in two ways. A record that a model created needs to say how it was made. Anyone who checks it later needs to be able to repeat the run.

The provenance record

The helper provenance() collects the facts from Ollama. workshop.llm.chat already logs every call (see below), so this cell shows the full provenance record just once. The cell writes the record with log_run, the same function the chat helper uses.

from workshop.llm import provenance, log_run
prov = provenance(MODEL)
record = {"notebook": "01", "prompt": messages[-1]["content"], "answer": response.message.content, **prov,
          "options": {"temperature": 0, "num_ctx": context_limit(MODEL)}}
log_run(record)
print({k: record[k] for k in ("model", "digest", "quantization", "ollama_version", "captured_at")})
print("appended to runs/log.jsonl")

{'model': 'nishad/apertus:8b', 'digest': '4a85a6fea0b8748229840a797cdeb5f45075b681040dc538f8ed17bf9e55059e', 'quantization': 'Q4_K_M', 'ollama_version': '0.40.4', 'captured_at': '2026-10-15T10:32:13+00:00'}
appended to runs/log.jsonl

From the executed notebook 01, 15 October 2026. Your digest should match this one if you downloaded the same tag. Your version and time will differ.

Notebook 00 shows the full record as JSON. Besides the fields above, it has parameter_size (8.05B), context_length (65536) and template_header.

The log file

runs/log.jsonl holds one JSON object per line, one per logged call. It is plain text: open it in an editor, load it with any JSON tool or search it with grep. Every workshop.llm.chat call appends a line with the time, the model and its digest, the options, the last 300 characters of the prompt and the first 300 of the answer. Streamed calls log the request only. The RAG helper, the tool loop and the setup check’s smoke calls (quick test calls) all go through chat, so they are logged too. Three kinds of call are not logged: embedding calls, the OpenAI client call in 01.06 and the oversized prompt in 04.07. The last two skip the helper on purpose. Keep the file with your results.

Citing a model in a workflow document

Name the model by tag and digest, not by tag alone. For example:

Records were generated with nishad/apertus:8b (digest 4a85a6fea0b8, Q4_K_M), run through Ollama 0.40.4 at temperature 0 with seed 42, context 8,192 tokens, on 15 October 2026. Prompts and answers are in runs/log.jsonl.

Add the full digest in an appendix or the log.

What a digest does and does not tell you

A digest identifies one thing: this exact file. It does not tell you what text the model was trained on. For that, you need the model’s public documentation, such as the EU training-data summary that Apertus publishes. The provenance chapter shows how to look it up. Record both: the digest for the file, the summary for the data.