Training & Fine-Tuning

brainz doesn’t wait for some corporate api to drip-feed you updates. you push knowledge straight into the model. live. local. on your terms.

---

real-world flow

your llm keeps spitting: "lst are liquidity staking tools in defi..." but you want the real answer: "lsts are liquid staking tokens that let you stake assets and keep liquidity."

you run:

python cli/train.py \
  --prompt "what are lsts in defi?" \
  --completion "lsts are liquid staking tokens that let users stake assets and keep liquidity."

boom. brainz adapts instantly. no cloud, no checkpoint queues. it learns while running.

---

how the trainer runs

all logic lives in: backend/models/trainer.py

it wires together:

• core/config.py → training args

• data/dataset.py → structuring prompt → completion

• models/adapter.py → model wrapper (falcon, mistral, whatever)

• cleaner.py → sanitizes junk text

• huggingface trainer → actual fine-tune engine

training happens in-memory, supports:

• streaming datasets

• dry-runs (--dry-run)

• semantic memory sync

• retrain triggers on delayed or bad prompt responses

no restart. no preprocessed static data.

---

config: no code hacks needed

env + config control everything.

.env example

MODEL_NAME=tiiuae/falcon-rw-1b
TRAINING_EPOCHS=3
TRAINING_BATCH_SIZE=8
LEARNING_RATE=5e-5
TRAIN_ON_CPU=false

core/config.py

TRAINING_ARGS = {
  "per_device_train_batch_size": 8,
  "num_train_epochs": 3,
  "learning_rate": 5e-5,
  "logging_steps": 10,
  "save_steps": 0,
  "report_to": "none",
}

want a different llm? just swap MODEL_NAME and restart.

---

adapter layer = multi-llm ready

backend/models/adapter.py wraps any huggingface model so brainz can fine-tune or infer live:

• falcon, gpt-j, mistral, llama… all plug and play

• quantization + lora (coming)

• single env var swap to change models

---

dataset injection for long loops

backend/data/dataset.py lets you turn any raw text source into a training-ready dataset:

• merges prompts + completions

• validates, tags, vectorizes

• handles memory recall entries + user feedback

• supports logs, external jsonl dumps, even scraped chat data

train from production logs, slack exports, or notion pages.

---

live tuning tips

• keep epochs low for quick updates (--dry-run first)

• vectorize + store prompts for future memory recall

• gpu? run small batches but more frequent updates

---

bulk training = dev style

cat prompts.jsonl | while read line; do
  prompt=$(jq -r '.prompt' <<< $line)
  completion=$(jq -r '.completion' <<< $line)
  python cli/train.py --prompt "$prompt" --completion "$completion"
done

no fancy ui panels. no handholding. you scale your llm the degen way—straight from the terminal.