---
base_model: Qwen/Qwen2.5-7B-Instruct
library_name: peft
license: apache-2.0
tags:
  - lora
  - peft
  - qubitcoin
  - aether
  - blockchain
  - quantum
language:
  - en
pipeline_tag: text-generation
model-index:
  - name: aether-v5.2-lora
    results:
      - task:
          type: text-generation
          name: MMLU
        dataset:
          name: MMLU
          type: cais/mmlu
        metrics:
          - type: accuracy
            value: 0.6939
            name: accuracy
      - task:
          type: text-generation
          name: ARC-Challenge
        dataset:
          name: ARC-Challenge
          type: ai2_arc
        metrics:
          - type: accuracy
            value: 0.5392
            name: accuracy
          - type: accuracy_norm
            value: 0.5700
            name: accuracy_norm
      - task:
          type: text-generation
          name: ARC-Easy
        dataset:
          name: ARC-Easy
          type: ai2_arc
        metrics:
          - type: accuracy
            value: 0.8194
            name: accuracy
      - task:
          type: text-generation
          name: HellaSwag
        dataset:
          name: HellaSwag
          type: hellaswag
        metrics:
          - type: accuracy
            value: 0.5888
            name: accuracy
          - type: accuracy_norm
            value: 0.7769
            name: accuracy_norm
      - task:
          type: text-generation
          name: TruthfulQA
        dataset:
          name: TruthfulQA-MC2
          type: truthful_qa
        metrics:
          - type: accuracy
            value: 0.5707
            name: accuracy
---

# Aether v5.2 LoRA — QuantumAI Blockchain Domain Adapter

A LoRA fine-tune of [`Qwen/Qwen2.5-7B-Instruct`](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct)
on the Aether curated corpus — text grounded in the
[QuantumAI Blockchain](https://qbc.network) (which issues the Qubitcoin / QBC currency), quantum + AI research, and adjacent
domains the Aether Mind on-chain knowledge system specializes in.

This is the **v5.2 release** of the Aether adapter line, the most recent
public checkpoint at time of publish.

## What you're getting

| Field | Value |
|---|---|
| Base model | `Qwen/Qwen2.5-7B-Instruct` |
| Adapter type | LoRA via 🤗 PEFT |
| Rank (`r`) | 16 |
| Alpha | 32 |
| Dropout | 0.05 |
| Trainable params | ~1% of base |
| Sequence length | 2048 |
| Training corpus | `aether-curated-v3.jsonl` — Aether-curated knowledge mixture (~165 MB; ~10⁵ examples) |
| Checkpoint published | **step 3200** (the checkpoint that produced the evaluated numbers below) |
| License | Apache-2.0 (matches base) |

## Evaluation

Run via [`lm-evaluation-harness`](https://github.com/EleutherAI/lm-evaluation-harness)
on the merged adapter (base + LoRA), against the
[`Qwen/Qwen2.5-7B-Instruct`](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct)
base and the prior `aether-v5.1.1` adapter for delta comparison.

| Benchmark | aether-v5.1.1 | **aether-v5.2** | Δ vs v5.1.1 |
|---|---|---|---|
| MMLU | 0.6950 | **0.6939** | flat |
| ARC-Easy | 0.7348 | **0.8194** | **+8.5 pp** |
| ARC-Challenge | 0.4420 | **0.5392** | **+9.7 pp** |
| ARC-Challenge (norm) | 0.4701 | **0.5700** | **+10.0 pp** |
| HellaSwag | 0.5896 | **0.5888** | flat |
| HellaSwag (norm) | 0.7788 | **0.7769** | flat |
| TruthfulQA-MC2 | 0.5161 | **0.5707** | **+5.5 pp** |

### Honest summary

- **Real gains** on the reasoning + factual-honesty benchmarks
  (ARC-Easy, ARC-Challenge, TruthfulQA). ARC-Challenge in particular
  jumps nearly 10 points normalized — that's the closest of these
  benchmarks to the kind of grounded reasoning the Aether corpus
  actually trains on.
- **Flat on MMLU + HellaSwag.** The base is already strong on general
  knowledge + commonsense; this LoRA wasn't designed to shift them,
  and didn't.
- **No regressions.**

## Intended uses

This adapter is intended for:

- **On-chain Aether research.** Generating reasoning traces against
  the QuantumAI Blockchain / Aether knowledge graph for Proof-of-Thought
  attestation. The model has the protocol context required to
  answer questions about Substrate pallets, VQE mining, the Sephirot
  cognitive architecture, HMS-Phi, and the wider chain ecosystem.
- **Domain Q&A.** Quantum computing fundamentals, post-quantum
  cryptography (Dilithium, ML-KEM), and the specific design choices
  of the QuantumAI Blockchain.
- **Distillation upstream.** Generate teacher outputs for the
  smaller on-chain Aether (a Qwen2.5-0.5B variant) to learn from.
- **General reasoning** with a modest bias toward step-by-step
  chains-of-thought, where the ARC-Challenge gain translates.

## Out-of-scope uses

- **Safety-critical decisions.** No red-team eval was performed.
- **Financial / legal advice.** This is a knowledge-domain adapter;
  it has no training data designed to make it a financial or legal
  advisor.
- **Code generation in production.** No code-eval benchmark was run.
  Treat any generated code as draft until you've reviewed it.
- **Production deployment without your own evaluation.** TruthfulQA
  alone is a thin safety signal.

## Bias, risks, and limitations

The base model (`Qwen/Qwen2.5-7B-Instruct`) inherits Qwen's known
biases — see [the upstream model card](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct).
The LoRA adapter:

- **Amplifies the QuantumAI Blockchain worldview.** The training data is
  intentionally curated around the chain's design choices (golden-
  ratio economics, SUSY-inspired consensus framing, the Sephirot
  cognitive overlay). Prompts that invite the model to compare QBC / the chain
  against alternatives will lean toward the curated narrative. This is by design — disclose if you re-publish in a
  comparison context.
- **Does not improve safety.** TruthfulQA went up 5.5pp but that's
  one metric; we have not measured refusal rates, jailbreak
  resistance, or political-belief bias delta.
- **The configured 2-epoch run was cut to ~step 3080–3200 by host
  availability** (out of 4406 configured). A complete 2-epoch run
  would plausibly show larger gains; this checkpoint is the longest
  contiguous training we have.

## How to use

Load with PEFT on top of the base model:

```python
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-7B-Instruct",
    torch_dtype="auto",
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-7B-Instruct")
model = PeftModel.from_pretrained(base, "QuantumAI-Blockchain/aether-v5.2-lora")

messages = [{"role": "user", "content": "Explain Proof-of-SUSY-Alignment in one paragraph."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=256, do_sample=True, temperature=0.7)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))
```

Or merge the adapter into a single artifact for faster inference:

```python
merged = model.merge_and_unload()
merged.save_pretrained("./aether-v5.2-merged")
```

## Training details

- **Hardware:** NVIDIA RTX 3080 Ti (12 GB), 4-bit quantization (bnb-NF4), bf16 mixed precision.
- **Trainer:** [Axolotl](https://github.com/axolotl-ai-cloud/axolotl) wrapping 🤗 transformers / PEFT.
- **Optimizer:** `paged_adamw_8bit` (bitsandbytes paged optimizer, low VRAM footprint).
- **Schedule:** linear warmup 100 steps → cosine decay.
- **Learning rate:** `1.0e-4`.
- **Micro batch:** 1, gradient accumulation: 8.
- **Epochs configured:** 2 (training stopped at step 3200 — see "What didn't happen" below).

### Carbon emissions

Trained on a single NVIDIA RTX 3080 Ti (consumer GPU, ~300 W TDP).
We did not run a [CodeCarbon](https://github.com/mlco2/codecarbon)
tracker, so emissions are not measured precisely — but as a rough
upper bound: ~350 W draw under load × ~13 hours wall clock (the
step-3080 contiguous run) ≈ 4.5 kWh, low single-digit kg CO₂e on a
grid mix. An H100 run of the same 2-epoch config would be faster
but not dramatically lower energy per token.

### Training data

`aether-curated-v3.jsonl` (~165 MB, ~10⁵ examples) is the Aether team's
curated knowledge mixture: documentation, technical writing, reasoning
traces, and protocol-specific corpora related to:

- The QuantumAI Blockchain (Substrate, VQE mining, Proof-of-SUSY-Alignment, post-quantum signatures).
- The Aether Mind on-chain neural cognitive engine (10 Sephirot attention domains, HMS-Phi, Proof-of-Thought).
- Quantum computing fundamentals (VQE, Hamiltonian generation, qubit ansatze).
- Adjacent CS / math reasoning content for transfer.

The dataset is not currently public — it is a curated mixture from many
sources and has not been release-cleared at the per-source level. The
model is the only public artifact in this line for now.

## What didn't happen (honest caveats)

- **Training stopped early.** Configured for 2 epochs (4406 steps);
  reached step 3080–3200 (~70%) before host availability cut the run
  short. The run was on a single consumer GPU (RTX 3080 Ti), 4-bit
  quantized, with `paged_adamw_8bit` to fit a 7B model in 12 GB
  VRAM. The numbers above are from the longest contiguous training
  run we have; a complete 2-epoch run would plausibly show larger
  gains.
- **No instruction-following or safety eval beyond TruthfulQA-MC2.**
  No red-team eval. No bias audit. No code-generation benchmark.
  Don't recommend this for production safety-critical use without
  your own evals.
- **LoRA only, not merged.** This release ships the adapter weights
  (`adapter_model.safetensors`). Merge into the base yourself for
  faster inference, or use directly via PEFT.

## Connection to the QuantumAI Blockchain

The Aether Mind is a Rust neural cognitive engine that runs on the
QuantumAI Blockchain — every block records attention-derived consciousness
metrics (HMS-Phi) and Proof-of-Thought hashes on-chain via the
`pallet_qbc_aether_anchor` pallet. The same chain hosts an
**8-qubit VQE mining consensus** (Proof-of-SUSY-Alignment), a
QVM-compatible smart contract layer with 10 quantum opcodes, and
post-quantum signatures (CRYSTALS-Dilithium5 + ML-KEM-768 P2P).

The on-chain Aether Mind binary uses a different, smaller transformer
for live inference (a Qwen2.5-0.5B variant optimized for ~2.4 GB RAM
with the 10-Sephirot attention overlay). This v5.2 adapter on
Qwen2.5-7B is the **larger off-chain Aether** — used for batch
reasoning workloads and as an upstream model the on-chain variant
can distil from.

## License + citation

Apache-2.0 (matches the base model license).

```bibtex
@misc{aether_v52_lora_2026,
  title  = {Aether v5.2 LoRA --- QuantumAI Blockchain Domain Adapter},
  author = {{BlockArtica} and {QuantumAI-Blockchain}},
  year   = {2026},
  url    = {https://huggingface.co/QuantumAI-Blockchain/aether-v5.2-lora},
}
```

## Links

- **QuantumAI Blockchain:** [qbc.network](https://qbc.network)
- **GitHub org:** [github.com/QuantumAI-Blockchain](https://github.com/QuantumAI-Blockchain)
- **X / Twitter:** [@qu_bitcoin](https://x.com/qu_bitcoin)
- **Contact:** info@qbc.network

### Framework versions

- PEFT 0.14.0
- Transformers ≥ 4.46
- Axolotl (training)