Fine-tuning recipe: all-MiniLM-L6-v2

python -m pip install "transformers>=4.40" "peft>=0.11" "accelerate>=0.30" "datasets>=2.19" "safetensors>=0.4"

from transformers import AutoModel
from peft import LoraConfig, TaskType, get_peft_model

base = AutoModel.from_pretrained(
    "sentence-transformers/all-MiniLM-L6-v2",
    revision="ea78891063587eb050ed4166b20062eaf978037c",
    trust_remote_code=False,
)

config = LoraConfig(
    task_type=TaskType.FEATURE_EXTRACTION,
    r=8,
    lora_alpha=16,
    lora_dropout=0.05,
    target_modules="all-linear",
)
model = get_peft_model(base, config)

# Train locally with your reviewed dataset and objective.
# FinanceGPT does not execute this training job.
model.save_pretrained("./financegpt-adapter", safe_serialization=True)

from transformers import AutoModel
from peft import PeftModel

base = AutoModel.from_pretrained(
    "sentence-transformers/all-MiniLM-L6-v2",
    revision="ea78891063587eb050ed4166b20062eaf978037c",
    trust_remote_code=False,
)
model = PeftModel.from_pretrained(base, "./financegpt-adapter")

# Optional local developer action; never automatic in FinanceGPT.
merged = model.merge_and_unload()
merged.save_pretrained("./merged-model", safe_serialization=True)

1. Keep the exact base repository and pinned revision in the adapter card.
2. Retain the base artifact SHA-256 when available.
3. Review dataset provenance and permissions before training.
4. Evaluate the adapted model before any publication request.
5. Publish explicitly; publication never promotes the model into production.
6. Prefer Safetensors for adapter and merged weight files.