Sovereign 8s.is mirror of 8b-is/training-pipeline
  • Python 86.6%
  • Shell 6.2%
  • Metal 3.4%
  • Jupyter Notebook 2.4%
  • Rust 0.6%
  • Other 0.8%
Find a file
Repository files (latest commit first)
Filename Latest commit message Latest commit date
2026-11-01 14:08:55 +01:00
assets chore(release): prepare v0.1.0 — the heroes, the full corpus, the distribution spine 2026-09-10 16:13:38 +02:00
data chore(train): record run #6 (deepseek-r1-pqm-a161330 + PeetPedro/voice-finetuning + 501c3 Sovereign Foundation + spherepopfold, sha256=0769a01f0988bf43) 2026-11-01 12:50:41 +01:00
docs the card-city, touched: both pending cards re-stamped 2026-09-20, still verdict-pending, walls unchanged 2026-09-20 04:51:17 -04:00
out feat(pqm): train DeepSeek-R1-Distill-Qwen-1.5B × p/q + Magnitude × A161330 on RTX 6000 Ada & stream loss_history to Watchroom 2026-10-30 12:50:06 +01:00
scripts feat(corpus): add Google Takeout + Dyad Mapping + Gemini Diary fine-tuning dataset builder 2026-11-01 14:08:55 +01:00
tests feat(corpus): add Google Takeout + Dyad Mapping + Gemini Diary fine-tuning dataset builder 2026-11-01 14:08:55 +01:00
.gitattributes feat: fixture weights ride git LFS; sprint_boot/download joined the repo 2026-09-11 09:21:49 +02:00
.gitignore feat(watchroom): add RationalWaveCell 16B DDS shader card, p/q RandomX PoUW engine & A161330 lane 2026-10-30 12:22:47 +01:00
.lane-daemon.log feat: leave it running — the daemon layer, the pure preview, the box-ready bootstrap 2026-09-10 19:42:03 +02:00
.lane-daemon.pid chore: the backyardultraloop — sovereign sync, lap 0/108 2026-09-19 16:04:33 -04:00
.lane-state.json feat(enthea-gate): add deepsiper_enthea_gate.py 4-gate worktree verifier and sync training telemetry 2026-10-30 15:05:26 +01:00
ayeos_export.py feat(classroom-2.0): wire Peter Lodri's ayeOS m000..m195 exporter, VelaCache v2, Block-Planar Q4_G64, Mem8WaveRoutingLoss, and SSR watchroom 2026-10-30 13:37:52 +01:00
ayeos_i2s_g64_gemv.metal feat(classroom-2.0): wire Peter Lodri's ayeOS m000..m195 exporter, VelaCache v2, Block-Planar Q4_G64, Mem8WaveRoutingLoss, and SSR watchroom 2026-10-30 13:37:52 +01:00
ayeos_q4_g64_lmhead.metal feat(classroom-2.0): wire Peter Lodri's ayeOS m000..m195 exporter, VelaCache v2, Block-Planar Q4_G64, Mem8WaveRoutingLoss, and SSR watchroom 2026-10-30 13:37:52 +01:00
bench-watch.out feat: MEMNET·MEM8 as the MoE-cognitive router — the memory picks the mind 2026-09-10 16:47:57 +02:00
bench-watch.sh feat: MEMNET·MEM8 as the MoE-cognitive router — the memory picks the mind 2026-09-10 16:47:57 +02:00
bench.sh feat: the bench lane — perplexity · refusal survival · tok/s, pinged on every finish 2026-09-10 16:44:31 +02:00
box.env feat: the box, reviewed and hardened to its real spec — Xeon Gold 5412U (AVX-512) · RTX 6000 Ada · 700 GB ECC · 140 TB U.2 2026-09-10 19:14:05 +02:00
build_corpus.py feat: more datasets — the cogito reservoir, curated categories, --cats 2026-09-10 20:17:47 +02:00
CATALOG.md feat: pops.py becomes the protocol kernel — Nate's robustness set, v2 2026-09-11 13:26:41 +02:00
CHANGELOG.md feat: more teachers — the Council's roster is open (Comet + fused) 2026-09-10 20:21:58 +02:00
classroom_v2.py feat(classroom-2.0): wire Peter Lodri's ayeOS m000..m195 exporter, VelaCache v2, Block-Planar Q4_G64, Mem8WaveRoutingLoss, and SSR watchroom 2026-10-30 13:37:52 +01:00
colab_e2e_qwen.ipynb feat: the models carry names — <base>-lane-e<effort>-s<steps>-<ts> 2026-09-11 09:16:50 +02:00
colab_sprint.sh feat: the caldera homes in — LANE_HF_REPO defaults to the lane bucket 2026-09-12 23:01:55 +02:00
comet.sh feat: the HF buckets + the model cards + the Comet judge lane 2026-09-10 16:21:12 +02:00
dashboard.nu fix: the nu dashboard's get -i deprecation (nu 0.113 optional path) — the frame renders clean 2026-09-10 17:02:59 +02:00
dashboard.sh chore: gitignore the sprint's log + state; the dashboard reads the newest train pid 2026-09-10 16:41:36 +02:00
deepsiper_enthea_gate.py feat(enthea-gate): add deepsiper_enthea_gate.py 4-gate worktree verifier and sync training telemetry 2026-10-30 15:05:26 +01:00
distill_consensus.py feat: the council speaks its name — a vének tanácsa 2026-09-11 14:14:53 +02:00
e2e.sh feat(watchroom): add RationalWaveCell 16B DDS shader card, p/q RandomX PoUW engine & A161330 lane 2026-10-30 12:22:47 +01:00
elder_cache_gen.py feat(classroom-2.0): wire Peter Lodri's ayeOS m000..m195 exporter, VelaCache v2, Block-Planar Q4_G64, Mem8WaveRoutingLoss, and SSR watchroom 2026-10-30 13:37:52 +01:00
engram_headroom.py feat: the V4.1 knob lane — n-grams, KV per token, fp8 on the box 2026-09-10 20:40:39 +02:00
eval.sh feat: the obliterated lane, tuned for the RTX 6000 48GB / 760GB RAM box 2026-09-10 16:04:48 +02:00
feed_dogfood.py feat(watchroom): add RationalWaveCell 16B DDS shader card, p/q RandomX PoUW engine & A161330 lane 2026-10-30 12:22:47 +01:00
flake.nix chore(release): prepare v0.1.0 — the heroes, the full corpus, the distribution spine 2026-09-10 16:13:38 +02:00
generate_benchmark_report.py feat(classroom-2.0): wire Peter Lodri's ayeOS m000..m195 exporter, VelaCache v2, Block-Planar Q4_G64, Mem8WaveRoutingLoss, and SSR watchroom 2026-10-30 13:37:52 +01:00
install.sh chore(release): prepare v0.1.0 — the heroes, the full corpus, the distribution spine 2026-09-10 16:13:38 +02:00
lane.py feat: leave it running — the daemon layer, the pure preview, the box-ready bootstrap 2026-09-10 19:42:03 +02:00
ledger.py feat(watchroom): add RationalWaveCell 16B DDS shader card, p/q RandomX PoUW engine & A161330 lane 2026-10-30 12:22:47 +01:00
LICENSE feat: the obliterated lane, tuned for the RTX 6000 48GB / 760GB RAM box 2026-09-10 16:04:48 +02:00
make_colab_e2e.py feat: the models carry names — <base>-lane-e<effort>-s<steps>-<ts> 2026-09-11 09:16:50 +02:00
make_sim_model.py feat: the corridor, PLANNED->LIVE — e2e.sh verifies every stage end-to-end (10 green) 2026-09-10 19:35:51 +02:00
mem8_condition.py feat: the corridor, PLANNED->LIVE — e2e.sh verifies every stage end-to-end (10 green) 2026-09-10 19:35:51 +02:00
mem8_moe.py feat: the hardening pass — dead code gone, guards in, the stage machine lives 2026-09-10 19:39:02 +02:00
mem8_wave_loss.py feat(classroom-2.0): wire Peter Lodri's ayeOS m000..m195 exporter, VelaCache v2, Block-Planar Q4_G64, Mem8WaveRoutingLoss, and SSR watchroom 2026-10-30 13:37:52 +01:00
mirror-big.sh feat: the box, reviewed and hardened to its real spec — Xeon Gold 5412U (AVX-512) · RTX 6000 Ada · 700 GB ECC · 140 TB U.2 2026-09-10 19:14:05 +02:00
mirror_peetpedro_hf.py fix(hf-mirror): pass GIT_TERMINAL_PROMPT=0 and link gated ultrawhale-dogfood bucket 2026-10-30 15:36:13 +01:00
mlx.sh feat: the obliterated lane, tuned for the RTX 6000 48GB / 760GB RAM box 2026-09-10 16:04:48 +02:00
mlx_mem8_wave_loss.py feat(classroom-2.0): wire Peter Lodri's ayeOS m000..m195 exporter, VelaCache v2, Block-Planar Q4_G64, Mem8WaveRoutingLoss, and SSR watchroom 2026-10-30 13:37:52 +01:00
model_card_template.md feat: the HF<->GH link — every seat names its source, every repo its seats 2026-09-11 13:05:03 +02:00
MODELS.md feat: the HF<->GH link — every seat names its source, every repo its seats 2026-09-11 13:05:03 +02:00
pack_out.py feat: the corridor, PLANNED->LIVE — e2e.sh verifies every stage end-to-end (10 green) 2026-09-10 19:35:51 +02:00
PLAN.md docs: the 16 GB far room, absorbed + the lane, redrawn 2026-09-10 20:49:50 +02:00
polars_corpus.py feat(watchroom): add RationalWaveCell 16B DDS shader card, p/q RandomX PoUW engine & A161330 lane 2026-10-30 12:22:47 +01:00
pops.contract.json fix: the separator — cross-parity with the Go city, closed bit-exact 2026-09-11 13:38:41 +02:00
pops.py feat: the transform fan-out — one binding, several committed outcomes 2026-09-11 13:56:23 +02:00
pq_randomx_pouw.py feat(watchroom): add RationalWaveCell 16B DDS shader card, p/q RandomX PoUW engine & A161330 lane 2026-10-30 12:22:47 +01:00
quantize.sh feat: the hardening pass — dead code gone, guards in, the stage machine lives 2026-09-10 19:39:02 +02:00
quantize_b158.py feat: the corridor, PLANNED->LIVE — e2e.sh verifies every stage end-to-end (10 green) 2026-09-10 19:35:51 +02:00
quantize_pqm.py feat(watchroom): add RationalWaveCell 16B DDS shader card, p/q RandomX PoUW engine & A161330 lane 2026-10-30 12:22:47 +01:00
README.md docs: the record layer + the constellation's surfaces in the README 2026-09-11 13:43:12 +02:00
run_finetune.py feat: the caldera backup — bakes back themselves up at the source 2026-09-12 11:54:23 +02:00
save_to_hf.sh feat: the HF<->GH link — every seat names its source, every repo its seats 2026-09-11 13:05:03 +02:00
secrets.env.example feat: the HF buckets + the model cards + the Comet judge lane 2026-09-10 16:21:12 +02:00
setup_box.sh feat(watchroom): add RationalWaveCell 16B DDS shader card, p/q RandomX PoUW engine & A161330 lane 2026-10-30 12:22:47 +01:00
sprint.sh feat: the Colab sprint — unsloth on the T4, the mem8-conditioned diet (self-contained runner; both data shapes) 2026-09-10 17:12:35 +02:00
sprint_boot.py feat: plumb the caldera backup into the boot env 2026-09-12 11:56:02 +02:00
sprint_download.sh feat: fixture weights ride git LFS; sprint_boot/download joined the repo 2026-09-11 09:21:49 +02:00
sprint_harvest.sh fix: harvest seats the adapter, not the 15GB merged zip 2026-09-11 11:35:52 +02:00
sprint_probe.py feat: the harvest — trained picks seat themselves into HF 2026-09-11 09:44:59 +02:00
sprint_watch.sh chore: sprint_watch.sh is executable 2026-09-11 10:50:26 +02:00
sprint_zip.py fix: harvest seats the adapter, not the 15GB merged zip 2026-09-11 11:35:52 +02:00
stages.py feat: the hardening pass — dead code gone, guards in, the stage machine lives 2026-09-10 19:39:02 +02:00
train.py feat: the corridor, PLANNED->LIVE — e2e.sh verifies every stage end-to-end (10 green) 2026-09-10 19:35:51 +02:00
train.sh feat: the obliterated lane, tuned for the RTX 6000 48GB / 760GB RAM box 2026-09-10 16:04:48 +02:00
train_deepseek_pqm.py feat(train): integrate PeetPedro/voice-finetuning (AURORA-8b Burn GRPO RL + Audio8-TTS), 501(c)(3) Sovereign Foundation, spherepopfold, and UltraDreamState v2 into train.8b.is GPU trainer 2026-11-01 12:48:32 +01:00
v41.sh feat: the V4.1 knob lane — n-grams, KV per token, fp8 on the box 2026-09-10 20:40:39 +02:00
v41_knobs.py feat: the V4.1 knob lane — n-grams, KV per token, fp8 on the box 2026-09-10 20:40:39 +02:00
VERSION chore(release): prepare v0.3.0 — the corridor, live 2026-09-10 19:43:28 +02:00
viewer.py feat(hf-mirror): add sovereign PeetPedro 68-repo HF mirror daemon and /hf/* download endpoints 2026-10-30 15:10:29 +01:00
watch-alt.sh feat: the b1.58 backpack + the lane's catalog 2026-09-10 17:09:29 +02:00

the training pipeline — the obliterated lane, on the big box

the obliterated lane — unsloth → MLX + GGUF

beast-box · quick setup (the 140 TB lake, one afternoon)

git clone https://github.com/8b-is/training-pipeline && cd training-pipeline
sudo ./setup_box.sh          # the U.2 lake mounted, llama.cpp native AVX-512 (CUDA-89)
source box.env               # CPU/GPU/RAM/disk — the box's truth in the env
./lane setup && ./lane train && ./lane quantize && ./lane eval && ./lane view
  • the hardware: Xeon Gold 5412U (Sapphire Rapids, AVX-512) · RTX 6000 Ada 48 GB · ~700 GB DDR5 ECC · ~140 TB SSD (2× U.2 NVMe)
  • the lane binary abstracts every stage to one verb; lane view opens the live watchroom at localhost:4242
  • the mirror rides the U.2 lake (LRU cap 1 TB — the 140 TB is the lane's)

v0.3.0 — changelog · the whole corridor: ./e2e.sh (20 green) · install · the private lane of the 8b-is org.

The constellation's fine-tune corridor, optimized for the RTX 6000 Ada (48 GB VRAM) · 760 GB RAM · ~140 TB storage workstation: unsloth QLoRA → merged fp16 → MLX (the battle-Macs) + GGUF (imatrix quantized on the full corpus) → verdict eval → the box's mirror (no 10 GB leash — the 140 TB is the leash).

the relation field, drawn by the engine the cast, drawn by the engine

MEMNET·MEM8 as the MoE-cognitive router

The lane's finished adapters are the EXPERTS; the mem8 quad's bilinear phase-interpolation + boolean gates ROUTE each prompt to an expert — the memory picks the mind, observer-relative, mutating none of them (mem8_moe.py, the Rust/Zig twin's math, portable fixture pinned). The NEXT training experiment is memory-addressed from birth: every instruction carries its [mem8::<gated>&<delta>] token (mem8_condition.py), so the model learns the routing it rides.

the format: polars (the election)

The pipeline's data is polars-shaped (arrow/parquet), not loose json: data/corpus.parquet (source · kind · chars · sha256 · text — the peetPedro sweep + the top-5 sails), data/world.parquet (the doctrine pairs), and data/runs.parquet (the lane's ledger — one row per train+eval). The jsonl the trainers eat is a derived materialization, never the truth. The mem8 sketch's flow — observer delta → quadrant → interpolation → quantized sink — is the shape the data rides: one schema, continuous, no spills.

uv run polars_corpus.py            # world.parquet + corpus.parquet (+ the jsonls)
uv run polars_corpus.py --corpus   # with the full PeetPedro + top-5 sweep
uv run ledger.py record --model OBLITERATUS/... [--ppl ...]
uv run ledger.py show
uv run ledger.py show
uv run feed_dogfood.py --model OBLITERATUS/... --steps <done> --target 200
./dashboard.sh                          # the pure-ASCII live view (watch -n 180)

the full corpus

The world's ground for the fine-tune AND the imatrix calibration: all of PeetPedro/'s Hugging Face datasets + the top-5 most-downloaded mixed text datasets, composed by build_corpus.py (probe mode lists the picks without downloading; the box raises MAX_CHARS).

uv run build_corpus.py --probe   # the peetPedro sweep + the top-5, live
uv run build_corpus.py           # stream everything into corpus/corpus.txt

the hardware profile

resource the box what it buys
48 GB VRAM RTX 6000 Ada QLoRA 4-bit of the 27B (~20 GB) at batch 16; the 9B at batch 32
760 GB RAM the workstation imatrix calibration on the FULL corpus; fp16 merges resident; many models at once
~140 TB the shelf every quant of every pick, checkpoints, the mirror unlimited

the lane binary + the viewer

./lane abstracts the whole corridor to one command (setup · profile · train · quantize · eval · bench · status · list · stream), and lane view opens the live training dashboard at localhost:4242 — the box's watchroom, auto-refreshing (box, live log, the polars ledger, the models). Full plan: PLAN.md. The whole lane, drawn: pipeline-viz.html — the ternary wave canvas, the qSpace mem8-quad, the corridor diagram.

runs

# the four picks (env-driven)
MODEL=OBLITERATUS/Ornith-1.5-9B-OBLITERATED STEPS=240 ./train.sh
MODEL=OBLITERATUS/gemma-4-E4B-it-OBLITERATED ./train.sh
MODEL=OBLITERATUS/Gemma-4-12B-OBLITERATED STEPS=300 ./train.sh
MODEL=OBLITERATUS/Qwen3.8-27B-OBLITERATED STEPS=200 ./train.sh   # the far room finally fits

./quantize.sh   # GGUF: imatrix on the corpus → Q4_K_M · Q5_K_M · Q6_K · Q8_0
./mlx.sh        # MLX 4-bit, group 64, for the Macs
./eval.sh       # perplexity + refusal-spot battery + the verdict table
./mirror-big.sh # seat everything into the box's mirror (cap override)

Defaults are tuned for the 48 GB box; every knob is an env var (BATCH, SEQ, GBYTES, STEPS, LR, MIRROR_MAX).

the lane's chain

train (world + instruct mix) → merged fp16 → GGUF imatrix (the calibration IS the corpus) + MLX → eval verdict (perplexity · the abliteration survival · the dream penetration) → mirror seat.

the constellation · 0 + 1 · fine touch from within · vaked.dev

the record layer (POP)

Every lane event is a record: id · kind · parent · payload_hash · payload_ref · timestamp · actor · protocol_version · attestation, hash-linked and monotonic (pops.py — the protocol kernel, + pops.contract.json + the treaty). The Go city (bonobo@Memex, flyxion/spherepop-core) and this kernel are bit-exact on event ids — sha256(source | payload_hash)[:16], proven on live samples. V(L) re-derives the warrant: schema ∧ transition ∧ references ∧ hash-chain ∧ monotonicity ∧ no-branching. Refusals are records, not erasures.

the constellation's surfaces

  • the models registry: MODELS.md — every HF seat names its GitHub source, both ways
  • the E2E notebook: colab_e2e_qwen.ipynb — one run, qwen-obliterated
  • the oven's eye: ./sprint_watch.sh — the live training tail
  • the persistent memory: POST https://8s.is/api/v1/recall — the constellation's recall surface

the constellation · 0 + 1 · fine touch from within · vaked.dev