88 lines
2.9 KiB
Markdown
88 lines
2.9 KiB
Markdown
# 2026-04-29 FO-Blog v7 / TIP Handoff
|
|
|
|
## Summary
|
|
|
|
Rene asked where the model from the last days of training is located. The screenshot from TIP showed `fo-blog-v7` active with:
|
|
|
|
```text
|
|
BLOG_LLM_PROVIDER=ollama
|
|
OLLAMA_LLM_MODEL=fo-blog-v7
|
|
```
|
|
|
|
Codex verified that the current active TIP BlogLLM is indeed `fo-blog-v7`, but it is not stored as a normal Ollama GGUF model named `fo-blog-v7`.
|
|
|
|
## Finding
|
|
|
|
`fo-blog-v7` is a virtual/local model name served by the Mac Studio Adapter Bridge.
|
|
|
|
Bridge definition:
|
|
|
|
```text
|
|
/Users/renefichtmueller/Desktop/Claude Code/magatama/scripts/ollama_adapter_bridge.py
|
|
```
|
|
|
|
Actual trained adapter:
|
|
|
|
```text
|
|
/Users/renefichtmueller/Desktop/Claude Code/magatama/training-data/runpod/pod-runs/2026-04-25-fo-tip/final/adapters/fo_blogllm/final-adapter
|
|
```
|
|
|
|
Important adapter file:
|
|
|
|
```text
|
|
adapter_model.safetensors
|
|
```
|
|
|
|
Observed details:
|
|
|
|
- Adapter directory size: about `169M`.
|
|
- `adapter_model.safetensors`: about `161.5M`.
|
|
- File timestamp: `2026-04-26 08:18`.
|
|
- Base model in bridge: `Qwen/Qwen2.5-7B-Instruct`.
|
|
- Bridge model name: `fo-blog-v7`.
|
|
- Aliases include `fo-blog-v7:latest`, `fo-blog-v6`, `fo-blog-v6:latest`, `fo-blog-v6-runpod`, `fo-blog-v6-runpod:latest`.
|
|
|
|
## TIP Wiring
|
|
|
|
TIP API default model:
|
|
|
|
```text
|
|
/Users/renefichtmueller/Desktop/Claude Code/github-repos/transceiver-db/packages/api/src/llm/client.ts
|
|
```
|
|
|
|
Relevant behavior:
|
|
|
|
```text
|
|
const OLLAMA_URL = process.env.OLLAMA_URL || "http://localhost:11434";
|
|
const LLM_MODEL = process.env.OLLAMA_LLM_MODEL || "fo-blog-v7";
|
|
const BLOG_LLM_PROVIDER = process.env.BLOG_LLM_PROVIDER || "ollama";
|
|
```
|
|
|
|
TIP dashboard card:
|
|
|
|
```text
|
|
/Users/renefichtmueller/Desktop/Claude Code/github-repos/transceiver-db/packages/dashboard/index.html
|
|
```
|
|
|
|
The dashboard correctly labels the model as `fo-blog-v7` and "Adapter Bridge / Mac Studio".
|
|
|
|
## Important Clarification
|
|
|
|
Do not search for `fo-blog-v7.gguf` as the canonical artifact. Codex checked the local Ollama manifests and found visible manifests for older models such as `fo-blog-v5`, `fo-blog-v6`, and `fo-blog-v3-qwen7b`, but not a normal `fo-blog-v7` manifest.
|
|
|
|
The correct artifact for the current `fo-blog-v7` is the PEFT adapter under the Magatama RunPod folder above.
|
|
|
|
## Decision
|
|
|
|
- Treat `fo-blog-v7` as the current active TIP BlogLLM model.
|
|
- Treat `fo-blog-v8` as the next model/training candidate.
|
|
- Keep the TIP crawler/robot planning lane TIPLLM-only.
|
|
- Keep Erik cautious: controller/light `erik-safe` only.
|
|
- Put heavy crawler execution on Proxmox/Pi workers.
|
|
|
|
## Open Points
|
|
|
|
- Verify the Adapter Bridge process health before a production blog-generation run.
|
|
- Decide whether to package/export `fo-blog-v7` into a GGUF/Ollama manifest for easier discovery, or keep the PEFT adapter bridge as canonical.
|
|
- When training `fo-blog-v8`, preserve the same handoff pattern: final adapter path, base model, bridge mapping, TIP env/default, and dashboard state must all be written into `sync/`.
|