PEFT LoRA adapter that specializes LiquidAI/LFM2.5-1.2B-Instruct for retention workflows (trained via Unsloth hub id unsloth/LFM2.5-1.2B-Instruct).
Creator: d4rkninja
Collection: Tanpo — Domain Specialists
Tanpo is a family of compact domain-specialized business models for local / edge / inexpensive deployment. Different specialists cover different workflows.
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).
Full eval, examples, and responsible-use notes: d4rkninja/tanpo-retention.
Best For
- Onboarding, activation, churn-save, renewal, expansion, escalation, and winback workflows
- Customer-success playbooks and account action-plan drafts
Not Designed For
- Automated customer decisions, guaranteed retention outcomes, or unauthorized system access
- General coding or non-retention chat
Adapter settings (from adapter_config.json)
| Field |
Value |
| Base (adapter_config) |
unsloth/LFM2.5-1.2B-Instruct (Unsloth-compatible of LiquidAI/LFM2.5-1.2B-Instruct) |
| PEFT type |
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) |
Evaluation (summary)
| Model |
Rubric overall |
| Base |
94.2% |
| Fine-tune |
96.0% |
| Delta |
+1.8 percentage points |
BEATS_BASE: YES. Strong on escalation, renewals, and winback at 100%; identity_guardrails is +34.3 pp vs base. Trails base on cs_playbooks (−22.2 pp) and health_scores (−5.5 pp).
Internal DarkLab automated domain evaluation (directional). Same prompts and generation config for base vs fine-tune. Not an industry benchmark. See the merged card evaluation/ for artifacts and methodology.
Load with PEFT
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = "unsloth/LFM2.5-1.2B-Instruct" # Unsloth-compatible of LiquidAI/LFM2.5-1.2B-Instruct
tokenizer = AutoTokenizer.from_pretrained(base, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
base, torch_dtype="auto", device_map="auto", trust_remote_code=True
)
model = PeftModel.from_pretrained(model, "d4rkninja/tanpo-retention-LoRA")
Related
Limitations
This is a compact domain adapter; verify outputs against account context and source systems. It is not a general-purpose or frontier model, and the automated evaluation is directional. The cs_playbooks and health_scores categories trail the base in the reported evaluation.
Responsible Use
Humans must review customer communications, retention offers, escalation plans, and consequential account decisions. Do not use for deception, harassment, discriminatory treatment, or unauthorized access.
License
This adapter repo is currently tagged license: apache-2.0 on Hugging Face. Using the adapter with the base model remains subject to the upstream Liquid AI LFM2.5 / LFM license on LiquidAI/LFM2.5-1.2B-Instruct (training hub id unsloth/LFM2.5-1.2B-Instruct). Treat base-model terms as controlling for combined use and redistribution of merged weights.
Tanpo Model Family
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.
Original upstream: LiquidAI/LFM2.5-1.2B-Instruct (~1.17B, 32K context)
└── Unsloth-compatible hub load: unsloth/LFM2.5-1.2B-Instruct
└── Tanpo Specialist (domain SFT via LoRA)
├── Full / Merged → d4rkninja/tanpo-retention
├── LoRA adapter → d4rkninja/tanpo-retention-LoRA
└── GGUF quants → d4rkninja/tanpo-retention-GGUF
