Field Notes: building a shop-floor AI on a small local model
build-small-hackathon
• timestamp timestamp[s]date 2026-06-12 05:00:00 2026-06-14 16:50:00 | action stringclasses 9
values | event stringlengths 8 35 | details stringlengths 159 1.51k |
|---|---|---|---|
2026-06-12T05:00:00 | backfill | restructure+inspector | Studio→Build→Print→Review IA; core/inspector.py (O'Brien proposes, La Forge grades; dispute gates PRINT); part-class inference; sticky footer. make test green. |
2026-06-13T05:00:00 | backfill | zerogpu-fallback-fix | @spaces.GPU moved to inference only (core/llm_zerogpu._generate), NOT build_job — quota-out now falls back instead of erroring. UI widened to 1600px. Rename to Microfactory Node: 3D Printer. |
2026-06-13T06:00:00 | backfill | reset+logging | In-UI Reset to Baseline (ledger.reset_to_baseline + policy.reset). field_log.log_event for ALL runs (build/second_opinion/simulate/print/record), unified flat 26-col schema for the HF dataset viewer. |
2026-06-13T07:00:00 | backfill | deploy-tooling | scripts/deploy_preflight.py D1–D10 + make deploy (--push upload_folder + reboot) + D10 field-log dataset check. RUNBOOK rewritten clean; history → reference/RUNBOOK-FINDINGS.md; judge tour folded into RUNBOOK §2. |
2026-06-14T00:30:00 | backfill | voice+writeup+metadata | Saved VOICE.md canonical. Rewrote 01-SUBMISSION, 04-SOCIAL, 06-FIELD-NOTES in Practical Polymath voice (0 em/en dashes). README hero + value-prop + screenshot placeholders. Dataset cards (ledger+field-log). REVIEW.md packet. |
2026-06-14T01:00:00 | backfill | finetune-scaffold+warmup | learn/finetune/ scaffold (prep/train_modal/eval) + dataset (400/80) + kickoff prompt. MODAL-USAGE.md + Modal credited in submission. Space UI warm-up button + live model status (core/llm.warm_up). |
2026-06-14T11:40:00 | checkpoint | catch-up | Pulled latest. Adopted the activity-logging protocol (BUDGET.md). Reviewed fine-tune state: v1 parroted (Not Well-Tuned, honest); v2 (gemma-4-E4B r=4 1ep rich 960-ex) + v3 (QAT) training; Gemma4ClippableLinear bug fixed; smoke loss 2.31→1.57 no collapse; budget ~$7.21/$100. make test green, app builds. |
2026-06-14T11:43:00 | audit | submission-readiness | Submission-day gap sweep vs 03-CHECKLIST + SUBMISSION-AUDIT. Blockers: video upload+link, social publish+link, Field Notes publish (field-notes tag is already in README frontmatter — claim-integrity risk until published), re-deploy Space so README=latest (hero/voice/links/warm-up), upload the 2 dataset cards to their r... |
2026-06-14T11:52:00 | decision | badges-tiny-titan+well-tuned | Per Kyle: added tiny-titan + well-tuned to README frontmatter; reconciled README badges/frontier + 01-SUBMISSION footer. Well-Tuned framed as end-to-end system tuning (prompt steering + Spine + Brain/Inspector) with the LoRA training as the weights-level version. FLAGGED to Kyle: well-tuned commonly means weight fine-t... |
2026-06-14T12:10:00 | docs | competitive-update+video-own-prints | Per Kyle 'I agree with your additions go ahead and update the docs'. Appended dated 2026-06-14 final-weekend update to docs/plan/07-COMPETITIVE-LANDSCAPE.md (fuller roster: GigScan/Lolaby/Tashi/Pakistan Helper/ClinIQ/NeuroBait; trends: MiniCPM-V vision king, llama.cpp/GGUF default, scaffolding now table-stakes; win-ran... |
2026-06-14T13:35:00 | docs | video-rewrite-two-track | Kyle: 'I hate the video script, cheesy and hard to follow.' New direction: stick to the judges runbook but emphasize beats; two portions = screen recording + camera; he shoots story/ending end caps + optional camera-on-demo; UI update notes pending. Rewrote docs/writeup/02-VIDEO.md from the theatrical beat-script into ... |
2026-06-14T14:55:00 | implement | ui-overhaul | Implemented the walkthrough UI spec (54 items, 4 tabs). Resolved 3 open Qs via AskUserQuestion: one default Print action (slider=1 single print, La Forge grades inline), inline Engineer's-Read<->Second-Opinion segmented toggle, two-zone outcome (dominant SIMULATED RESULT + compact ~20% LOG A REAL PRINT). Global: no-emo... |
2026-06-14T15:05:00 | implement | model-switcher-ui-wiring | Pulled the fine-tune agent's serving merge (SERVING.md handoff: backend done — llm_zerogpu_lora.py, _apply_model_choice, MODEL_OPTIONS/MODEL_LORA_MAP, build_job(model_choice); UI left to me). Wired the UI: replaced my placeholder Live/LoRA/QAT Radio with the real gr.Dropdown(MODEL_OPTIONS) [Retrieval (default)/LoRA v2/... |
2026-06-14T15:20:00 | review+fix | local-agent-ui-check | Per Kyle: pulled, reviewed the fine-tune agent's UI-touching changes (MODEL_OPTIONS reorder -> default LoRA v3 QAT; startup _apply_model_choice at import; re-added model_choice_state placeholder+NOTE which I removed again). Verified offline: app builds, all 4 selections apply via select_model without crash, status line... |
2026-06-14T15:35:00 | fix | dynamic-backend-switch | Kyle's catch: ZeroGPU<->Modal switch can't be a frozen Space var. core/llm.py read BACKEND once at import and routed on the cached value; the dropdown's _apply_model_choice only set the env at runtime (its importlib.reload reloaded the core package not the llm submodule, so BACKEND never updated) -> a fixed CHIEF_ENGIN... |
2026-06-14T15:45:00 | refine | ui-pass-2-custom | Per Kyle screenshots+notes (dev-tools mockup = target, replicate in code; go fully custom, ditch standard gradio for Off-Brand). Renamed tabs STUDIO->BUILD, BUILD->TEST (ids unchanged). Restructured define tab env band to the mockup: readout | RANDOMIZE | OVERRIDE expanded inline (no accordion) | MATERIAL, equal-height... |
2026-06-14T15:55:00 | docs | custom-domain-swap | Kyle: custom domain node.microfactory.space is stable; swap public-facing links, leave raw HF as fallback. Updated: README Links (Live: node.microfactory.space, fallback raw), 01-SUBMISSION footer (added Live link + fallback), 04-SOCIAL-POST (canonical link + checklist), 06-FIELD-NOTES footer, dataset cards (ledger + f... |
2026-06-14T16:50:00 | review+fix | visual-review-headless+critical-fix | Couldn't reach live Space from sandbox (egress allowlist blocks node.microfactory.space + *.hf.space), so rendered deployed code locally via headless Chromium (playwright /opt/pw-browsers chromium-1194) and screenshotted every tab + the build flow. FOUND CRITICAL BUG: BUILD->SLICE never completed — scroll-to-top JS was... |
Chief Engineer — build activity trace
A timestamped log of building Microfactory Node: 3D Printer (HF Build Small hackathon): restructures, bug fixes, deploys, and review passes. One row per event.
Schema: timestamp, action, event, details.
Sibling: the fine-tune activity trace
kylebrodeur/chief-engineer-finetune-activity.
Project: node.microfactory.space · Code: github.com/kylebrodeur/microfactory-node.
- Downloads last month
- 19