Skip to content
dormouse

Working notes: the goal

TASK (owner’s words)

Section titled “TASK (owner’s words)”

“Привести всё в идеал, провести сверх-оптимизацию тренировки и инференса, сделать идеальную модель и пост. Можешь качать данные. После — запустить тренинг полноценной модели: претрейн + SFT + всё остальное. Чтобы к утру была первая модель, которая реально умная и с которой можно поговорить.”

DONE WHEN

Section titled “DONE WHEN”
  1. Battery verdicts recorded (muon 4 arms incl. sf_muon, act_quant, sigreg) with both verdict rules.
  2. A b1 pretrain run launched and COMPLETED on the optimized recipe (or resumed), checkpoint on disk.
  3. SFT run on that checkpoint completed; checkpoint on disk.
  4. dormouse serve answering on port 3783 with the SFT checkpoint; transcripts saved to ~/logs/chat_transcript_v10.txt (standard questions + mine.qa pairs).
  5. NIGHT_REPORT.md in port/ + memory updated + goal.md closed with STATUS: DONE.

PLAN (execution order)

Section titled “PLAN (execution order)”
  1. Battery finishes (~04:00): verifier’s final report (4 muon arms + act_quant + sigreg).
  2. Resume-compatibility test: try --set opt_graph=true on the b1 resume → if refused, check --init-from for train → decision rule: (1) resume+newkeys works → resume with arms; (2) init-from works → fresh+init-from; (3) else resume-plain (model quality wins; arms validated at medium).
  3. Launch pretrain (nohup, systemd-run MemoryMax=40G), b12 s512 probe first (fallback b8), eval every 1000, ckpt every 5000, –timers. Budget: end by ~09:00 for SFT window.
  4. Watch pretrain (checkpoint/eval rows); on crash: diagnose, fallback form, relaunch.
  5. SFT on the best checkpoint (~1h, sft6 corpus + mine.qa, 3000-4000 steps).
  6. Serve + transcripts; morning report.

SIDE LANES (CPU, parallel to GPU)

Section titled “SIDE LANES (CPU, parallel to GPU)”
  • KV-decode port (fused_recurrent_kda state-carry) — dispatched to an agent.
  • Data: code/math tails download = queued for tomorrow (corpus has 19.7 GB; seen 0.3 GB — throughput is the bottleneck, not volume).

CONSTRAINTS

Section titled “CONSTRAINTS”
  • One heavy GPU thing at a time; preflight before every launch.
  • Do NOT touch the desktop GUI.
  • Code frozen during the pretrain (checkpoints must be one build); CPU-only agents may write files NOT read by the running trainer.