lora-qwen25-1p5b-finbot-v2

FinBot LoRA — Qwen2.5-1.5B-Instruct (v2)

A LoRA adapter that fine-tunes Qwen2.5-1.5B-Instruct to act as the recommendation engine of FinBot, a multilingual (English / Vietnamese / Chinese) financial-planning assistant designed for fully local, privacy-first deployment.

The adapter specialises the base model for one job: given structured user context collected by a deterministic dialogue engine, produce a schema-valid JSON recommendation in the user's language — reliably, concisely, and safely.

• 📦 GitHub (full system): https://github.com/vutuongvy101/financialbot
• 👩‍💻 Author: Tuong Vy Vu — GitHub · LinkedIn

Why fine-tune?

The base 1.5B model could not reliably emit valid structured output for this task (0% schema validity on the held-out evaluation set). Rather than scaling up the model or adding heavy post-processing, this adapter was trained via teacher–student distillation to fix the failure mode directly at the model level.

Evaluation

Results on a held-out evaluation set, base model vs. base + this adapter:

Evaluated on 62 held-out examples balanced across English (21), Vietnamese (21), and Chinese (20), spanning planning, investment, and trading tasks. Latency measured per-example on an NVIDIA A100-80GB. Schema validity reached 100% in every language individually, not just in aggregate. Evaluation covers cascading JSON schema validation, safety/refusal behaviour (unsafe-claim screening, e.g. "guaranteed profit"), structured internal-analysis presence, and output efficiency.

Training

• Method: Supervised fine-tuning (LoRA/PEFT, via TRL SFTTrainer) on teacher–student distilled examples, filtered and validated against the FinBot output schema before training.
• Data: 615 validated multilingual instruction–response pairs (from 618 raw across 9 source files; 2 removed by the multilingual safety filter, 1 deduplicated, 0 schema failures). Balanced across languages — EN 208 / VI 208 / ZH 199 — and tasks — planning 210 / trading 209 / investment 196. Split 553 train / 62 eval (90/10, seed 42). No real user data was used; all training data is synthetic.
• LoRA configuration:
  • rank (r): 16, alpha: 32, dropout: 0.05
  • target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj (attention + MLP)
  • trainable parameters: 18.5M of 1.56B (1.18%)
• Training run: NVIDIA A100-80GB (Google Colab), bf16 LoRA — no quantization or gradient checkpointing needed at this VRAM. 2 epochs, learning rate 2e-4 (cosine schedule, 3% warmup), effective batch size 6, max sequence length 2048 with sequence packing. The runbook is VRAM-adaptive and falls back to QLoRA (4-bit NF4) with gradient checkpointing on <24 GB GPUs.
• Reproducibility: seed 42 throughout; full run provenance (hyperparameters, dataset stats, metrics, library versions) recorded in the adapter's manifest.json.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "Qwen/Qwen2.5-1.5B-Instruct"
adapter_id = "bibbbu/lora-qwen25-1p5b-finbot-v2"

tokenizer = AutoTokenizer.from_pretrained(base_id)
model = AutoModelForCausalLM.from_pretrained(base_id, device_map="auto")
model = PeftModel.from_pretrained(model, adapter_id)

messages = [
    {"role": "system", "content": "You are FinBot, a financial planning assistant. Respond with a single JSON object matching the FinBot recommendation schema."},
    {"role": "user", "content": "Profile: age 30, income bucket B, risk tolerance: moderate, goal: retirement savings, horizon: 25 years, language: en. Generate a recommendation."},
]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)
output = model.generate(inputs, max_new_tokens=512)
print(tokenizer.decode(output[0][inputs.shape[-1]:], skip_special_tokens=True))

The adapter runs on CUDA and Apple Silicon (MPS). For the full system — deterministic dialogue state machine, slot parsing, multilingual i18n, PII bucketing, prompt-injection defence, and JSON validation/repair pipeline — see the FinBot repository.

Intended use & limitations

Intended use: research and demonstration of structured-output fine-tuning for small, locally deployed LLMs; portfolio/educational use; as the recommendation component inside the FinBot system, where its output is schema-validated before display.

Limitations:

• ⚠️ Not financial advice. This model is a research prototype. Its outputs must not be used to make real investment, trading, or financial-planning decisions.
• Trained entirely on synthetic, distilled data — it inherits the biases and blind spots of the teacher models and the prompt distribution used for generation.
• A 1.5B-parameter model can and will hallucinate; the adapter improves format reliability and conciseness, not factual financial knowledge.
• The model expects inputs in the FinBot prompt format (pre-bucketed user context). Free-form financial Q&A is out of distribution.
• Coverage is limited to English, Vietnamese, and Chinese.

Citation

@misc{vu2026finbotlora,
  title  = {FinBot LoRA: Structured-Output Fine-Tuning of Qwen2.5-1.5B for a Multilingual Local Financial Assistant},
  author = {Vu, Tuong Vy},
  year   = {2026},
  url    = {https://huggingface.co/bibbbu/lora-qwen25-1p5b-finbot-v2}
}