HSJ/ lab
HSJ Lab
READ ONLY
0.8B · RAG · HYDROPONICS

Hydroponics knowledge lab

Trusted sources are organized into a searchable knowledge base (RAG). A small model learns to answer from those sources and cite them, and independent evaluations check the results. Every step is recorded here.

Current phase

Preparing (training not started)

Not scheduled · Run manually stage by stage

Loading status…
—New sources
—Accepted knowledge units
—Quarantined knowledge units
Not startedFine-tuning in this run
00

Research progress

  1. Done
    Training environment

    GPU passthrough and driver verified: RTX 3050 Laptop 4 GB, 3,768 MiB of usable VRAM.

  2. Done
    Model test

    Qwen3.5-0.8B loaded, LoRA attached and one training step run: peak VRAM about 2,700 MiB; hand-calculated parameter counts match the actual model.

  3. Done
    Knowledge sources

    17 sources approved: 11 from Taiwan’s agricultural research stations, 4 from FAO and US university extension services, and 2 of my own notes.

  4. Done
    Knowledge base and retrieval

    Organized into 758 searchable passages; in a 23-question retrieval check, the correct passage was in the top 5 for 21 questions.

  5. Done
    Trial run and baseline preview

    A 100-step LoRA trial with five reference passages per example (about 2,400 tokens): peak VRAM about 2,900 MiB, 47 minutes. With the knowledge base, the untrained model answered 23 questions but often misread passages and answered anyway when the sources were insufficient — exactly what fine-tuning should fix.

  6. Done
    Building the test set

    289 questions: about 40 each on EC, pH, plant nutrition, plant physiology, hydroponic techniques and plant pathology, plus 50 general questions; every question was reviewed by the site owner, and 33 use the owner's revised answers.

  7. Done
    Knowledge base frozen

    Added 7 owner-verified nutrient-solution tables and a pH measurement note: 772 searchable passages; the retrieval check found the correct passage in the top 5 for 21 of 23 questions.

  8. Done
    Baseline evaluation

    Untrained model (automatic scoring): 25% of domain questions passed without sources and 52% with the knowledge base; 72% of general questions passed. Weakest points: declining when sources are insufficient, and overly long answers.

  9. In progress
    Generate training data

    Build question + reference passages + answer training examples from the knowledge base, each reviewed by the site owner.

Peak GPU memoryN/A
Peak RAMN/A
GPU temperatureN/A

Observation log

No training events yet. Live logs will appear here once training starts and the worker is connected; see “Research progress” above for the current status.

Read-only viewerLast heartbeat: not received

Research results do not directly control hydroponics equipment. Training execution and capability gains are recorded separately; promotion requires independent evaluation.