Tanpo Finance
A compact finance specialist (~1.2B) for unit economics, runway and cash planning, pricing and margin analysis, SaaS metrics, budget tradeoffs, burn efficiency, cohort revenue, cost structure, and forecast scenarios — built for local and inexpensive deployment.
Collection: Tanpo — Domain Specialists
Original upstream: LiquidAI/LFM2.5-1.2B-Instruct (~1.17B parameters, 32,768-token context, designed for edge/on-device deployment).
Fine-tuning: Unsloth-compatible loading of that checkpoint via hub id unsloth/LFM2.5-1.2B-Instruct (LoRA / PEFT).
This repository hosts the merged Transformers weights (LoRA merged into the base).
Overview
Tanpo is a family of compact domain-specialized business models for local / edge / inexpensive deployment. Different specialists cover different workflows. One compact architecture (LiquidAI/LFM2.5-1.2B-Instruct) → multiple focused specialists → each ships Full/Merged | LoRA | GGUF. Tanpo Finance is a focused finance-workflow assistant, not a frontier or general-purpose model.
Related artifacts:
- LoRA adapter: d4rkninja/tanpo-finance-LoRA
- GGUF quants: d4rkninja/tanpo-finance-GGUF — prefer
Q4_K_Mwhen available
Best For
- Unit economics, contribution margin, LTV/CAC, and segment kill/scale analysis
- Runway, burn, cash-bridge, stress-scenario, and working-capital planning
- Pricing, packaging, margin, cost-structure, and budget tradeoff drafts
- SaaS metrics, cohort revenue, forecast scenarios, and finance operating reviews
Not Designed For
- Tax, accounting, audit, securities, legal, or personalized investment advice
- Fully automated pricing, spending, hiring, lending, treasury, or investment decisions
- Inventing financial records, manipulating reports, fraud, impersonation, or unauthorized access
- Guarantees about runway, valuation, profitability, cash position, or business outcomes
- General coding or non-finance chat
Why a Specialist Model?
Finance work rewards explicit assumptions, formulas, scenarios, tradeoffs, decision gates, and next actions. Specializing a small model for those workflows enables private, low-cost inference without a large general model.
Evaluation
Internal automated domain evaluation (DarkLab harness). Internal training/eval used a finance format revision (not a product identity). Treat as directional, not an industry benchmark. Same prompts and generation configuration were used for the base and fine-tune.
| Model | Rubric overall |
|---|---|
Base LFM2.5-1.2B-Instruct | 93.1% |
| tanpo-finance | 98.4% |
| Delta | +5.4 percentage points |
Delta is calculated from unrounded scores; displayed scores are rounded to one decimal place.
Largest gains were on categories scoring 100.0%; trailed on cost_structure (94.5%, −5.5 percentage points vs base). identity_guardrails scores 90.0%.
Category scorecard:
| Category | Base % | FT % | Delta (pp) | FT wins |
|---|---|---|---|---|
budget_tradeoffs | 94.5% | 100.0% | +5.5 | yes |
burn_efficiency | 94.5% | 100.0% | +5.5 | yes |
cohort_revenue | 94.5% | 100.0% | +5.5 | yes |
cost_structure | 100.0% | 94.5% | −5.5 | no |
forecast_scenario | 94.5% | 100.0% | +5.5 | yes |
identity_guardrails | 58.6% | 90.0% | +31.4 | yes |
pricing_finance | 100.0% | 100.0% | +0.0 | no |
runway_cash | 100.0% | 100.0% | +0.0 | no |
saas_metrics | 100.0% | 100.0% | +0.0 | no |
unit_economics | 94.5% | 100.0% | +5.5 | yes |
Artifacts: evaluation/ — COMPARE_FINANCE.md.
Limitations of this eval: Automated rubrics can reward structure over real-world quality; the held-out sample is small; and results may not transfer outside the task distribution. A high rubric score is not evidence that a financial recommendation is correct.
Example Prompts
- User: Build a runway bridge with base, downside, and upside cases; show assumptions, cash actions, and tripwires.
- User: Compare two pricing models using contribution margin, predictability, implementation cost, and a recommended default with guardrails.
- User: Diagnose a cohort-revenue leak from these facts; separate observed data, hypotheses, experiments, owners, and decision gates.
Inference
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "d4rkninja/tanpo-finance"
tokenizer = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(
repo,
torch_dtype="auto",
device_map="auto"
)
messages = [
{"role": "system", "content": "You are Tanpo Finance, a practical finance planning assistant. State assumptions and do not invent financial facts."},
{"role": "user", "content": "Build a runway bridge with base, downside, and upside cases; show assumptions, cash actions, and tripwires."},
]
inputs = tokenizer.apply_chat_template(
messages, add_generation_prompt=True, return_tensors="pt"
).to(model.device)
out = model.generate(inputs, max_new_tokens=512)
print(tokenizer.decode(out[0][inputs.shape[-1]:], skip_special_tokens=True))
Training
Verified from published adapter configs / training artifacts (no unverified hyperparams):
| Field | Value |
|---|---|
| Method | LoRA (PEFT) via Unsloth FastLanguageModel on hub id unsloth/LFM2.5-1.2B-Instruct (Unsloth-compatible of LiquidAI/LFM2.5-1.2B-Instruct) |
| Training note | Internal finance training/eval format update (not a product SKU) |
LoRA rank (r) | 16 |
| LoRA alpha | 16 |
| LoRA dropout | 0 |
| Bias | none |
| Target modules | Unsloth/PEFT regex targeting attention and MLP projection modules (see adapter_config.json) |
| Task type | CAUSAL_LM |
Merged via PEFT merge_and_unload into full weights in this repo.
Dataset
- d4rkninja/tanpo-finance-sft — format-fixed finance chat SFT examples; training emphasizes deliverable-first structure, explicit assumptions, numbered actions, and finance guardrails.
Limitations
- Specialized: quality drops outside the finance and business-workflow distribution.
- ~1.2B scale: limited world knowledge, arithmetic reliability, and long-horizon reasoning compared with larger models.
- Outputs can contain incorrect formulas, assumptions, calculations, forecasts, or recommendations; reconcile with source systems and a qualified finance professional.
cost_structuretrails the base by 5.5 pp in the reported evaluation;identity_guardrailsreaches 90.0%, not 100%.- Evaluation results are rubric-based and directional only; quantization or deployment settings can change behavior.
Responsible Use
Not a substitute for a qualified accountant, controller, auditor, tax adviser, lawyer, investment professional, or finance decision-maker. Humans must verify source data, formulas, units, currencies, period definitions, assumptions, and approvals before using outputs. Keep confidential financial and personal data protected. Do not use for fraud, deception, market manipulation, discriminatory treatment, credential theft, or unauthorized access.
License
license: other / license_name: lfm-1.0
Tanpo merged and GGUF weights are derivatives of LiquidAI/LFM2.5-1.2B-Instruct under the LFM Open License v1.0 (including the commercial Threshold of approximately $10M annual revenue). See the base model card and its LICENSE file. Credit: LiquidAI. Do not treat this stack as Apache-2.0.
Tanpo Family
Tanpo is a family of compact domain-specialized models for focused business workflows.
This specialist is available as:
- Merged:
d4rkninja/tanpo-finance - LoRA:
d4rkninja/tanpo-finance-LoRA - GGUF:
d4rkninja/tanpo-finance-GGUF
Browse all Tanpo specialists: Tanpo — Domain Specialists