aether-v7.2-unified

The Sephirot adapter served by qbc.network, and the first adapter produced by our decentralised training loop to pass a full evaluation and be promoted into the live served model. It succeeds aether-v7.1-unified.

It is a small adapter over a frozen Qwen2.5-7B-Instruct base (Q4_K_M): ten low-rank experts with a top-2 router, about 1.18 million trainable parameters. The base model is not redistributed here; load Qwen/Qwen2.5-7B-Instruct and apply this adapter on top. Serving precision is BF16.

On-chain attestation

This model's identity is attested on the QBC chain. The on-chain weight_root is a SHA-256 over (base, adapter, tokenizer, config):

• weight_root: 0x3e99b3d2f4e4d5eacbb3fe8d7e2f709dbfc48deb34407035383b3a5263e29e4e
• registry round: 2, finalised by a 4-of-5 validator quorum
• submitted at block: 1698355
• content root of the trained adapter: 0x157a01b9

How it was trained

Trained with a DiLoCo-style local-training-plus-aggregation loop across a fleet of ordinary 8GB consumer GPUs. A vocab-chunked backward pass keeps a full 7B adapter step inside roughly 7.5GB of video memory, so commodity cards can train the served model. Training runs as short payable rounds of about 35 minutes: a worker draws a fresh shuffled batch from a large corpus, trains the adapter locally, and submits a candidate. A candidate is accepted only if it beats the previous accepted candidate on a fixed held-out set it never trains on. Ten rounds produced seven accepted units, and the best by full held-out evaluation is this release.

Evaluation

Measured on this adapter over the frozen base, as a snapshot from the promotion evaluation (not a live probe):

Measurement	Base	This adapter	Change
Held-out cross-entropy (clean 1000-example corpus, nats)	2.9014	2.5609	-0.3405
MMLU (full 14042, uncapped)	0.7116	0.7109	-0.07pp (flat)
GSM8K (full 1319, uncapped)	0.7248	0.7498	+2.50pp

The BF16 weights served here reproduce the held-out CE of the F32 candidate to four decimals, so the -0.3405 nats generalisation gain holds at serving precision. Generalisation on unseen text improved, general knowledge held flat, and multi-step maths reasoning improved, at no measured cost to base capability. Against the previously-served adapter, the promotion gate measured a further held-out CE improvement (2.5912 to 2.4811 on the gate's own holdout).

Scope and honesty

This adapter is the product of decentralised training with an owner-gated promotion into the served model. This is Stage 1 of a four-stage plan. It is not yet a permissionless model update: that stage is designed, not built, and is gated on robust aggregation that survives a malicious contributor, a stake-and-slash economic layer, and honest independent re-derivation of the claimed improvement.

During this work our own evaluation flattered us three times, through memorisation on training data, a sampling-window bug, and a mixed-scale gate that false-rejected a good round. Each was caught and fixed before any number here was trusted. The full account is in the engineering write-up.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-7B-Instruct")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct")
model = PeftModel.from_pretrained(base, "QuantumAI-Blockchain/aether-v7.2-unified")

Licence

Released under Apache-2.0, matching the Qwen2.5-7B-Instruct base licence.