Muse-Glimmer-30B-FinanceIND

An Indian personal-finance assistant fine-tuned from Meta's open Muse Glimmer 30B. It helps with SIPs, tax planning (old vs new regime, 80C, 80D, HRA), loans and EMIs, emergency funds and goal planning, calls calculation tools, and retries when a tool call fails.

What it can do

• Greets users, explains what it can do, and asks clarifying questions when details are missing
• Personalizes advice when a user profile is provided in the system prompt (or asks for one when it isn't)
• Uses tools for exact calculations and recovers from tool errors
• Declines tax evasion and guaranteed-return requests, and says when it has no live market data

Training

• Method: QLoRA (r=32, alpha=32), 1 epoch, no truncation, then merged into BF16 weights
• Data: Gandalf1/indian-finance-synthetic-phase2-cleaned (synthetic, quality-filtered), plus generated conversation, identity and tool-failure/retry examples
• Loss on assistant turns only

Usage

The BF16 weights are ~55 GB. On a 24-40 GB GPU, load in 4-bit:

from unsloth import FastModel
model, tok = FastModel.from_pretrained("Rumiii/Muse-Glimmer-30B-FinanceIND", max_seq_length=8192, load_in_4bit=True)
FastModel.for_inference(model)

Recommended sampling: temperature 1.0, top_p 0.95, top_k 64.

Limitations

• Tax rules in the training data follow FY 2024-25. India's Income-tax Act, 2025 applies from 1 April 2026, so always verify current rules and figures.
• Educational guidance only, not regulated investment or tax advice.
• It has no live market data and does not remember users between chats unless the app passes their profile.
• Synthetic training data can contain errors; check important numbers.

License: Apache 2.0 (same as the base model).