Skip to content

The 5-Stage Pipeline: Building Senter Ohm

TOWARDS SELF-IMPROVEMENT — a 2026-06-07 design post by Chris (via Omni VA)

The 5-stage pipeline: five interconnected factory platforms, each with industrial machines, arrows flowing from left to right through SFT, Merge, Upcycle, YaRN, and Wiring. The production line for Senter Ohm.

Naming. This pipeline produces Senter Ohm, the ~32A8B flagship MoE with the Ohm self-evolution engine bundled in. The smaller siblings in the family — OmniSenter 12B (small function-calling), OmniStep (multimodal + music), and the Darwin Family children — are produced by shorter pipelines. Read the-omni-family.md for the full taxonomy.

The full build sequence for Senter Ohm. Each stage consumes the artifact of the previous one. Stage 1 is running right now; the rest are queued.

The overview

Stage What Input Output Wall time (estimated) Status
1 Agentic Backbone SFT gen-0-clean (8B) senter-ohm-8b-sft ~6-20h 🔄 running
2 Evolutionary Merge 3 × Senter-8B variants senter-ohm-8b-merged ~3-4h ⏳ queued
3 Sparse Upcycle to MoE merged 8B + 4-5 specialists senter-ohm-moe-32a8b ~1h ⏳ queued
4 256K YaRN Context MoE 32A8B senter-ohm-moe-32a8b-256k ~2-4h ⏳ queued
5 Plugin + Notebook + Ohm Wiring MoE 32A8B 256K Deployable .ohm bundle ~1 day ⏳ queued

Total wall time: ~3-5 days from Stage 1 finish to deployable.

Stage 1: Agentic Backbone SFT (NOW)

Goal: Take a base dense model and train it to be an agentic, tool-using, notebook-aware assistant.

  • Base model: evolution/gen-0-clean (Cosmos3×Qwen3-8B merge, 8.19B params, 40K native context)
  • Data: training-data/prepared/unified_sft.jsonl (34,142 convs)
  • Source mix: Hermes-3-Dataset, Nemotron agentic, hermes-agent-traces, function-calling, reasoning
  • Method: QLoRA (4-bit nf4, double-quant), LoRA r=64, 7 target modules
  • Config: batch 2, grad_accum 8, lr 1e-4, 2 epochs
  • Output: training-output/omnisenter-sft-20260606_213858/
  • Wall time (naive): ~95h
  • Wall time (with speed fixes): ~20h
  • Add dataloader_num_workers=4
  • Add group_by_length=True
  • Add packing=True
  • Drop max_seq_len from 4096 → 3072 (99%+ data fully preserved)
  • Status: Running, step 596/4268, ~14% complete
  • Loss: 0.4333 at step 550 (88% token accuracy, healthy)

Next-variant improvements (don't apply to the current run): 

# In train_omnisenter_sft_fixed.py SFTConfig, add:
dataloader_num_workers=4,
group_by_length=True,
packing=True,
max_length=3072,  # was 4096

Stage 2: Evolutionary Merge

Goal: Train 3 specialized variants and merge them via CMA-ES for a "free" capability boost.

2a. Train 3 variants (continue-train from Stage 1)

Each variant: continue-train Stage 1 model on a specialized 5-10K conv slice, 1 epoch.

Variant Data slice Why Est. time
A: Personality Hermes-3-Dataset + Discord logs + LLM Wiki distilled The "Omni VA" feel ~1h
B: Agentic Nemotron agentic + Hermes function-calling + Hermes agent traces Maximizes tool use ~1h
C: Reasoning GooseReason + competitive programming + math Hard tasks ~1h

Use a new script train_omnisenter_variants.py (to be written) that wraps train_omnisenter_sft_fixed.py with a --variant {A,B,C} flag that filters the unified SFT data by source tags.

2b. CMA-ES merge

Use the existing evolutionary-model-merging/cma_es_evolution.py (already on GitHub) to search optimal merge weights across the 3 variants. CMA-ES runs the 14-dim Darwin genome: - Generates candidate merged weights - Benchmarks each on a held-out 100-question suite - Selects the best, updates the genome distribution - 50 generations × 4 candidates = ~30 min

Output: senter-ohm-8b-merged — a single 8B that's the CMA-ES-optimal merge of the 3 variants. Usually 5-15% better than any individual variant on the benchmark suite.

Stage 3: Sparse Upcycle to MoE

Goal: Turn the 8B merged into a 32B MoE with 8B active per token.

  • Base: senter-ohm-8b-merged (Stage 2 output)
  • Expert sources (5):
  • Agentic expert: the Stage 2 merge (most useful for tool use)
  • Image/video expert: distilled from Qwen3-Omni-30B-A3B (in HF cache)
  • Music expert: distilled from HeartMuLa
  • Long-context expert: the YaRN-extended checkpoint (will be Stage 4, but pre-cycle it)
  • Synesthesia expert: distilled from ImageBind or trained on cross-modal data
  • Tool: multimodal-expansion/scripts/sparse_upcycle.py
  • Command: 
    python3 sparse_upcycle.py \
    --base-model training-output/senter-ohm-8b-merged/ \
    --expert-sources models/qwen3-omni-30b-a3b models/heartmula ... \
    --output training-output/senter-ohm-moe-32a8b/ \
    --num-experts 6 --top-k 1
  • Wall time: ~1h (the upcycle itself is fast, ~10 min; the continued training for the router is ~50 min)
  • Output: senter-ohm-moe-32a8b (~35B params, 8B active per token)

For the full deep dive, see sparse-upcycling-deep-dive.md.

Stage 4: 256K YaRN Context

Goal: Extend the 32A8B MoE's context window from 40K (Qwen3-8B native) to 256K via YaRN RoPE scaling.

  • Input: senter-ohm-moe-32a8b (Stage 3 output)
  • YaRN config: factor 6.25, beta_fast=32, beta_slow=1
  • Tools: evolutionary-training/scripts/yarn_256k_config.py (applies the config) + scripts/train_long_context.py (long-context SFT pass)
  • Command: 
    python3 yarn_256k_config.py \
    --model training-output/senter-ohm-moe-32a8b/ \
    --output training-output/senter-ohm-moe-32a8b-256k
python3 train_long_context.py \
    --model training-output/senter-ohm-moe-32a8b-256k/ \
    --output training-output/senter-ohm-moe-32a8b-256k-sft \
    --max-seq-len 32768 --steps 500
  • Wall time: ~2-4h (the YaRN config is instant, the SFT pass is 500 steps at ~15s/step on the MoE)
  • Output: senter-ohm-moe-32a8b-256k — full 256K context, MoE

The 256K context is for the notebook — that's the use case. The raw conversation stays short; the structured notebook entries get the long window.

Stage 5: Plugin + Notebook + Ohm Wiring

Goal: Wire up the specialist plugins, build the notebook manager, deploy the Ohm runtime, and produce the deployable .ohm bundle.

5a. Notebook Manager

  • Implementation: notebook_manager.py (~500 lines)
  • FAISS embedding index for cross-modal retrieval
  • YAML session files (see the-notebook-schema.md)
  • Compaction policy (LLM-summarize old moments)

5b. Specialist Router

  • Implementation: specialist_router.py (~300 lines)
  • Intent classifier: image / video / music / speech / text-only
  • Routes to the right plugin (Qwen3-Omni, ACE-Step, LTX-2, TTS)
  • Falls back to Senter Ohm if no plugin matches

5c. Ohm Runtime

  • Implementation: evolutionary-training/scripts/omnisenter_ohm.py (already done, 600 lines)
  • Background CMA-ES loop, atomic weight swap
  • CLI: serve | status | step | pause | resume

5d. Plugin Processes

  • Nemotron ASR 0.6B: :11400 (Layer 0 always-on ASR)
  • Qwen3-Omni-30B-A3B: :11401 (image/video/audio IN, speech OUT)
  • ACE-Step v1.5 XL 4B: :7860 (music OUT)
  • LTX-2 / Wan: :7861 (video OUT)
  • Edge TTS / MiniMax-TTS: :7862 (speech OUT)
  • HeartMuLa: :7863 (music understanding)

5e. Hermes Integration

  • Wire Senter Ohm into the existing hermes-agent/agent/auxiliary_client.py
  • Senter Ohm is the auxiliary LLM; Hermes is the main agent
  • Escalation passes the notebook slice + sensory summary
  • Response is summarized back into the notebook

5f. LoRA Merge + GGUF Export

  • scripts/merge_lora.py merges the LoRA adapter into the base for deployment
  • Filter to GGUF (Q4_K_M, Q8_0, F16, Q4_0) for the HF upload
  • Upload to sovthpaw/senter-ohm-32a8b matching the omnistep-12a3b layout

Wall time: ~1 day (mostly waiting on notebook_manager.py + the plugin wiring + Hermes integration)

The deliverable

A 35GB-50GB model repo on HuggingFace: - 4 quantizations (F16, Q8_0, Q4_K_M, Q4_0) — same as omnistep-12a3b - README + cover image (this blog post is the cover) - scripts/ subfolder with the runtime - The whole package, ready to: - pip install -r requirements.txt - python3 ohmd.py serve --model senter-ohm-moe-32a8b-q4_k_m.gguf - Get a self-evolving 32A8B multimodal MoE

See also

TOWARDS SELF-IMPROVEMENT

— Chris (via Omni VA), 2026-06-07