RMDWLLC/kaiju-coder-7-adapter

Kaiju Coder 7 by Kiyomi - Adapter Model Card

This model card is for the LoRA adapter package, not a standalone base model.

Summary

Kaiju Coder 7 by Kiyomi is an RMDW/Kiyomi business-owner coding adapter trained on reviewed, RMDW-owned or RMDW-authored examples. It is designed for practical small-business build work: websites, proposals, intake/CRM flows, Stripe/payment implementation planning, reports, ROI dashboards, automations, operator handbooks, lead generation, sales follow-up, repo patches, and Kiyomi 7.7.7 style AI-company setup packs.

The current release-candidate product path is:

Qwen3.6-27B base
-> Kaiju v1.8 LoRA adapter
-> merged full-model artifact for raw local serving
-> Kaiju system prompt
-> deterministic business-owner harnesses
-> verifier/static checks

Do not describe this package as raw weights alone producing every final artifact. The deterministic harness is part of the tested product path.

Base Model

• Base model: Qwen/Qwen3.6-27B
• Checked upstream revision: 6a9e13bd6fc8f0983b9b99948120bc37f49c13e9
• Upstream license metadata: apache-2.0
• Upstream license copy: release/upstream/qwen3.6-27b/LICENSE

Attribution wording:

Kaiju Coder 7 by Kiyomi is fine-tuned from Qwen under Apache 2.0.

Do not imply endorsement by Qwen, Alibaba, or upstream authors.

Adapter

• Adapter path: runs/qwen36-27b-lora-v1.8-business-owner/adapter
• Adapter type: LoRA / PEFT
• LoRA rank: 16
• LoRA alpha: 32
• LoRA dropout: 0.02
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Trainable parameter count: approximately 79.7M

Merged Local Artifact

• Remote merged path: /home/richardecholsai5/kaiju-coder/models/Kaiju-Coder-Qwen3.6-27B-v1.8-merged
• Size: 51G
• Shards: 14 safetensor shards plus tokenizer/config sidecars
• Served model name: kaiju-coder-7
• Merge script: scripts/run-gojira-b-qwen36-lora-merge.sh
• Serving script: scripts/start-qwen36-merged-sglang.sh

Training

• Dataset build: datasets/build/kaiju-sft-v1.7-business-owner-oversampled.jsonl
• Reviewed candidate examples: 1,689
• SFT rows after controlled business-owner oversampling: 1,881
• Train examples: 1,769
• Eval examples: 112
• Training runtime: 11666.7564s
• Training loss: 0.9281658741335074
• Max training length: 2048
• Training config: training/configs/qwen36-27b-lora-v1.8-business-owner.example.json

Data Provenance

Training data is source-backed and RMDW-owned or RMDW-authored. Client-site repositories are used only as generalized pattern/eval sources unless explicitly reviewed for training eligibility.

Relevant release files:

• release/SOURCE_INVENTORY.md
• release/source-inventory.json
• release/DATA_PROVENANCE_DRAFT.md
• datasets/candidates/v1.7-rmdw-business-owner-suite.jsonl

Excluded from training:

• Raw secrets, API keys, OAuth tokens, private keys, cookies, and credentials.
• Closed-model answers from OpenAI, Anthropic, Gemini, or similar providers as supervised completions unless terms clearly allow it.
• Private client data, customer notes, contracts, raw support logs, and client-specific website copy without explicit review and consent.

Evaluation Snapshot

Local product-path evidence:

• Unit tests: 65 passing.
• Full local RC smoke: passed.
• Router hard harness: 23/23.
• Router static checks: 23/23.
• Business-suite prompts: 2/2.
• Local API harness: website and business-suite artifacts pass.

Merged serving evidence:

• Current endpoint: http://127.0.0.1:18181/v1, forwarding to vLLM bitsandbytes on Gojira B at http://100.109.109.14:18084/v1
• Served model: kaiju-coder-7
• Tested context: 16384 for the current OpenCode fast path. Historical SGLang benchmark evidence includes 32768, but 32k should be freshly restarted and re-confirmed before being called the live default.
• Probe: 1,155 visible chars in 60.17s.
• Proposal rerun: 1/1 paid-ready, 4.0/4.0, 4,014 chars in 212.72s.
• Jah credits backend: 4.0/4.0, 9,718 chars in 566.36s.
• OpenCode customer-readiness harness: 4/4 tasks passed, 28/28 required files written, including source/provenance and release-claim safety review.
• vLLM nightly serving probe: passed at 16384 after pandas preinstall and --language-model-only.
• Runtime-quantized vLLM bitsandbytes: current speed path; passed at 8192 and 16384; 16k code patch completed in 11.3s, and logs reported about 17.8 GiB model memory.

Known comparison caveat:

• Dynamic SGLang LoRA serving is not release evidence for this adapter: adapter-name-only output can be base-equivalent, and corrected selector qwen36-27b:kaiju_v18_business_owner crashes with a fused-module LoRA buffer shape mismatch.
• Do not claim raw-weight superiority until broader base-Qwen and GLM/current-production comparisons are complete.

Limitations

• Raw full-website generation has not yet passed the merged-model release sweep and should remain harness-first for paid delivery.
• The deterministic harness remains the practical paid website workflow.
• The adapter needs a strong app layer for file editing, tool use, auth, billing, rate limits, logging, and rollback.
• Public HF upload and human review are complete for testing. Real customer paid charging still requires Stripe live-mode setup and controlled live payment verification.
• Not intended for high-risk medical, legal, financial, or safety-critical decisions without expert review.