Parcimonia LFM3-VL-2B LoRA
This repository contains the Parcimonia v3 PEFT/LoRA adapter trained on
LiquidAI/LFM3-VL-2B. It is an experimental, text-only adapter for structured
Parcimonia routing decisions. It is not the Parcimonia router software and does
not replace its deterministic checks or verification layer.
Intended use
Research and local evaluation of advisory routing outputs for Parcimonia. Treat all generated decisions as untrusted proposals. Validate them in the calling application; do not use this adapter to authorize or execute actions.
Training and evaluation data
The v3 training manifest records 256 synthetic training questions, including 96 new training pairs and 32 replay pairs. Its evaluation set records 64 questions (32 pairs), with 32 expected abstentions. No images were used. The data are synthetic and authored for this experiment; these results do not establish performance on representative production traffic or independently collected data.
Limitations
- Experimental adapter; not independently evaluated.
- Evaluation examples are synthetic and authored with the experiment.
- Text-only: this adapter has no vision adapter.
- Keep the host application's deterministic policy, evidence checks and abstention behavior in control.
- The PEFT config records the upstream base model. Load that base model before loading this adapter.
Load with Transformers and PEFT
from peft import PeftModel
from transformers import AutoModelForImageTextToText, AutoProcessor
base_id = "LiquidAI/LFM3-VL-2B"
adapter_id = "zedgamer/parcimonia-lfm3-vl-2b-lora"
processor = AutoProcessor.from_pretrained(base_id, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
base_id, trust_remote_code=True, device_map="auto"
)
model = PeftModel.from_pretrained(model, adapter_id, is_trainable=False)
Use the upstream model's supported Transformers and PEFT versions. This card does not claim that the adapter has been tested across runtimes or hardware.
License
The base model is distributed under the LFM Open License v1.0. That license must accompany redistribution of this adapter. Review its redistribution and commercial-use conditions before use. The adapter's fine-tuned weights are a derivative of that base; do not treat Parcimonia's software Apache-2.0 license as applying to these weights.
Provenance
- Adapter revision: Parcimonia v3 local training artifact.
- Base model:
LiquidAI/LFM3-VL-2B. - Training set: synthetic; no images.
- Adapter SHA-256 (source manifest):
aa497a2fd4c62a2af1979007c715266a9bf053f4d3a2ab8ea728a942458a024b. - The companion Parcimonia application is at https://github.com/zedarvates/parcimonia.
- No GGUF is included in this package.
Tasks studied and evaluation results
The v3 adapter is trained to explain reasons in structured reports generated by the deterministic Parcimonia compute advisor. The studied tasks are distinguishing resource snapshot age from benchmark age, distinguishing machine identity from the measured batch fingerprint, and identifying other rejection reasons for a specified route. The advisor determines the route or abstention; an external guard checks the proposed reason against the report and requires auto_act_allowed=false. The adapter does not perform these checks independently.
In a 64-question synthetic held-out comparison, v3 produced 40 exact explanations versus 33 for v2; 17/24 versus 11/24 on age, and 12/24 versus 9/24 on identity. Other control reasons declined from 13/16 to 11/16. The guarded pipeline accepted 64/64 outputs for v3 versus 56/64 for v2 and preserved 32/32 expected abstentions. These guard results concern the combined evaluation pipeline, not the model alone. This authored benchmark is narrow, and does not demonstrate cost, speed or energy savings or production readiness.