scripts/README.md

Scripts

These are the scripts that produced the runs, copied as they were run (comments in Italian). Paths are absolute paths on the author's machines (/mnt/mneme-nvme/proxy-ideas/alb-bench-v2, a Strata install under /mnt/mneme-nvme/strata-0140): edit them before rerunning.

What you need

• AgentLongBench code (commit 57e13e8 of https://github.com/euReKa025/AgentLongBench) and data (https://huggingface.co/datasets/ign1s/AgentLongBench), laid out as alb/src/AgentLongBench and alb/data/benchmark/.... Prompts and scorer are the official ones, imported unchanged by common.py.
• An OpenAI-compatible engine serving the model (here Strata 0.1.40 with Qwen3.8 Flash Next IQ3_XXS, KV int8, 10 experts), started once per window size with --max-context equal to the window.
• mnemonic-proxy at commit b8d4920 (0.3.5) for the proxy method (PROXY_SRC), configured as in proxy-cfg-D.json.
• The model's tokenizer (PACK_TOK), used to count prompt tokens exactly as the engine does.

Files

file	role
make_main_lists.py	writes the preregistered job lists for all 25 questions per type and length (100 questions), in main-64k.txt, main-128k.txt
make_clar.py	builds the clarified questions (alb-clar/, same as questions-clarified/) and main-clar-*.txt
main-*.txt, main-clar-*.txt	job lists: one line per job, <policy> <window> <file> <qid> (A = full window, D = proxy, B = summary, T = truncation)
run_alb_v2.py	runs one job: builds the official messages, applies the method, asks the question, scores it with the official scorer, writes result.json and the logs
common.py	official prompt/parser/scorer imports, token counter, HTTP client
meter.py	engine call counter (engine_calls.jsonl) and resource sampler (resources.jsonl)
normfmt.py	format-normalised score (secondary; the headline numbers use the raw official score)
check_c2.py	stop rules checked after every job
gpu_block_c2.sh, gpu_block_clar.sh	one GPU block: starts the engine per window, runs the jobs of a list in order, stops on a rule; _c2 was used for the main run, _clar (same, with ALB pointing at the clarified data) for the confirmation run
run_main.sh, run_clar.sh	drivers: one block per length (64k then 128k)
build_results.py	builds results/*.jsonl of this dataset from the result.json files (extract logs/*.tar.zst into raw/runs/ next to it first)

Rerunning

python make_main_lists.py            # or use the lists shipped here
python make_clar.py                  # confirmation run only
bash run_main.sh 64k 128k            # main run      -> runs/main-64k, runs/main-128k
bash run_clar.sh 64k 128k            # confirmation  -> runs/main-clar-64k, runs/main-clar-128k

Sampling is seeded (seed 1), but GPU inference is not bit-for-bit deterministic: expect run-to-run noise (see the dataset README, Limits).