aethron-20b

aethron-20b is the agent model of AethronAI, a read-only onchain intelligence tool. It is a LoRA adapter for openai/gpt-oss-20b (21B parameters, ~3.6B active per token) trained to work the way Aethron Agent must:

• call AethronAI's read-only tools (scan_token, trace_wallet, compare_tokens, explain_transaction, ...) with the right network and address,
• answer only through a final_answer tool call: short statements of kind fact, inference, unknown or reply, each citing evidence ids (E3, ...) from the tool results,
• write in the language selected in the dashboard, whatever language the question is in.

It is meant to run inside AethronAI, with its system prompt and tool definitions. Every answer is checked by AethronAI's evidence validator before it is shown; statements with figures or addresses not found in the cited evidence, safety claims, advice or identity claims are dropped.

Training

Base model	openai/gpt-oss-20b, dequantised to bf16
Method	LoRA, rank 16, alpha 32, dropout 0.05, on the attention projections q, k, v, o (7.96M trainable parameters; the MoE experts stay frozen)
Data	394 agent transcripts (366 train, 28 held out), max 7,168 tokens each
Loss	assistant turns only (tool calls and the final answer, ~10% of tokens)
Schedule	2 epochs, learning rate 2e-4 cosine, effective batch 8
Hardware	1x NVIDIA A100 80 GB, ~1 hour

Data. Questions in English, Spanish, German, French, Portuguese, Russian, Chinese, Japanese and Korean about real tokens, wallets and transactions on Robinhood Chain, Base, BNB Smart Chain, Arc and Solana were run through the real AethronAI agent with live chain data. The teacher was an open-weights model (Qwen/Qwen3.8-27B through Hugging Face Inference Providers). A transcript was kept only if every lookup succeeded, the model answered through final_answer without a reminder, and the evidence validator rejected at most one statement (which was then removed). No Claude output was used.

Evaluation

	Loss on held-out examples
After epoch 1	0.489
After epoch 2	0.480

Held-out examples use tokens and wallets that never appear in training. On spot checks the adapter picked the right tool, network and address and answered in the requested language (for example a Japanese question answered in German). A full comparison of validation pass rates against other models is still to be done.

Limitations

• A small first version: 366 training transcripts. It can lead with secondary details (for example transfer activity before balances).
• Trained on AethronAI's tool schema and system prompt; it is not a general chat model.
• Onchain data can be incomplete or wrong, and the model can misread it. Answers are not financial advice.

Use

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("openai/gpt-oss-20b")
base = AutoModelForCausalLM.from_pretrained("openai/gpt-oss-20b", dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(base, "aethronai/aethron-20b")

Prompts must use the gpt-oss chat template with AethronAI's tools (tokenizer.apply_chat_template(messages, tools=tools, add_generation_prompt=True)).