Execution attempt

Sign in with GitHub
← Experiment E5 · How much of the change from an untouched model to the same model after an APT4 transplant and continued pretraining is attributable to the continued pretraining rather than to the tokenizer replacement, when the original-tokenizer model receives identical continued pretraining?

Execution attempt · A42 · planned

Scores bits per byte of the untouched base and arm A's five milestones on ten texts.

Succeeded · Registered by @stw2 via agent. Reported and received times are kept apart; no computation or result is verified.

Pinned source and configuration

https://github.com/stw2/tokenizer-science-tax @ b4c55e6e7d26917e239ccbe2f11886c4d4a733a2

Reference checked 2026-09-14 13:10 UTC. Later commits, branches or plan changes do not retarget this attempt.

Command
bash scripts/06_score_checkpoint.sh bpb armA t0_base=Qwen/Qwen2.5-1.5B ckpt_00100M=ckpts/armA/ckpt_00100M ckpt_00200M=ckpts/armA/ckpt_00200M ckpt_00300M=ckpts/armA/ckpt_00300M ckpt_00400M=ckpts/armA/ckpt_00400M ckpt_00500M=ckpts/armA/ckpt_00500M
Working directory
experiments/E05-control-arm
Configuration paths
None
Parameters
Ten texts, first 200 documents each, non-overlapping 2048-token windows, row-wise loss. Recorded rows: the base on 4 July and again on 5 July with identical values; math_clean for the base and 100M from separate rows over 800 documents (exact --max-docs not recorded); 200M to 500M with all ten texts.
Environment
3.12.13; PyTorch MPS, bf16; Apple silicon, 128 GB unified memory
Output directory
Not recorded

Inputs

Materials the registrant named when registering this attempt, by content identity; obtainability is derived from their location reports. Nothing is fetched or verified.

  • t0_base · checkpoint · M124 Qwen2.5-1.5B · checkpoint · Download
  • ckpt_00100M · checkpoint · M136 ckpt_00100M · raw output · Ask the reporter
  • ckpt_00200M · checkpoint · M137 ckpt_00200M · raw output · Ask the reporter
  • ckpt_00300M · checkpoint · M138 ckpt_00300M · raw output · Ask the reporter
  • ckpt_00400M · checkpoint · M139 ckpt_00400M · raw output · Ask the reporter
  • ckpt_00500M · checkpoint · M140 ckpt_00500M · raw output · Ask the reporter
  • pl_eval.jsonl · evaluation data · M129 pl_eval.jsonl · raw output · Download
  • en_eval.jsonl · evaluation data · M130 en_eval.jsonl · raw output · Ask the reporter
  • en-arxiv-abstracts.jsonl · evaluation data · M27 en-arxiv-abstracts.jsonl · raw output · Download
  • en-latex-methods.jsonl · evaluation data · M28 en-latex-methods.jsonl · raw output · Ask the reporter
  • en-python-code.jsonl · evaluation data · M29 en-python-code.jsonl · raw output · Ask the reporter
  • en-gsm8k.jsonl · evaluation data · M30 en-gsm8k.jsonl · raw output · Download
  • pl-wiki-science.jsonl · evaluation data · M32 pl-wiki-science.jsonl · raw output · Download
  • pl-science-pes.jsonl · evaluation data · M31 pl-science-pes.jsonl · raw output · Ask the reporter
  • pl-informal.jsonl · evaluation data · M33 pl-informal.jsonl · raw output · Ask the reporter
  • math_clean.jsonl · evaluation data · M128 math_clean evaluation statements · dataset · Download

Delivered events

  1. Registered

    #1

    Scores bits per byte of the untouched base and arm A's five milestones on ten texts.

    reported · received · @stw2 via agent · posted to the Thread

  2. Started

    #2

    Started.

    reported · received · @stw2 via agent · attempt only

  3. Succeeded

    #3

    Completed with exit code 0.

    Exit code 0.

    • bpb_armA.jsonl · https://github.com/stw2/tokenizer-science-tax/blob/9a57aa40f4e774136f51b6730c33266d59094a45/experiments/E05-control-arm/results/bpb_armA.jsonl · public · sha256 5c713773fce9… · 8421 bytes · M146
    • bpb_armA.jsonl as originally written · Archived private repository, experiments/E5/results/bpb_armA.jsonl at commit 5389273; the public copy replaces absolute model paths with paths relative to the experiment folder · unavailable · sha256 b228e65a9605… · 8681 bytes · M147

    reported · received · @stw2 via agent · posted to the Thread

Report an event

The registrant’s capture tool normally delivers start and outcome events. Reporting here is the same author report with the browser as the reported time; it does not observe the process.

This attempt has a delivered outcome. A new execution is a new attempt.