Feval

Scaling reliable and verifiable LLM evaluation across decentralized environments through distributed, scalable evaluation.

[Feval @Feval47]

Decentralization should be proven, not claimed.
SUBNET 47 Feval is built around open participation, independent validators, randomized audits, and verifiable results.
Our latest evaluation scored us 6/10 — and we welcome it.
It gives us a clear benchmark to improve from. As the subnet becomes fully operational, we aim to reach 10/10.
Transparency, reliability, and continuous improvement are what make a subnet trustworthy.
Build openly. Verify independently. Keep improving.
#Bittensor #SUBNET47 #Feval #DecentralizedAI

[Feval @Feval47]

So, how does Subnet 47- Feval actually work?

The process is simple:

1. Start with the same base model
Every miner begins from a fixed open-source model.

2. Improve it with LoRA
Miners use lightweight LoRA adapters to improve the model efficiently.

3. Generate responses
Each improved model runs against a shared evaluation set covering mathematics and instruction-following.

4. Score the results
Validators use transparent, deterministic rules to measure performance.

5. Verify with random audits
Selected outputs are checked through unpredictable GPU verification.

This creates a simple principle:

Improve → Evaluate → Verify → Reward

Cheap scoring keeps evaluation scalable, while targeted GPU verification helps keep results honest.

The best part?

Miners compete on measurable improvement, not just model size.

#Bittensor #Feval #AI #LLM #DeAI

[Feval @Feval47]

Hello everyone 👋

We’re the new founding team behind Subnet 47 — Feval.

Feval is a Bittensor subnet focused on one simple question:

Can we measurably improve an open-source AI model?

Miners start from a fixed base model and improve it using lightweight LoRA adapters.

Their models are then evaluated on a shared set of tasks covering areas such as:

→ Mathematics
→ Instruction following
→ General model performance

Validators score the results using transparent and deterministic rules, while randomly selected outputs are further verified through unpredictable audits.

The goal is not to reward bigger models.

It’s to reward genuine, measurable, and verifiable improvement.

We’re excited to build Feval with the Bittensor community.

#Bittensor #AI #LLM #DeAI