muse-qwen3.8-2b-tools-v7

Muse Qwen3.8 2B Tools v7

Muse Qwen3.8 2B Tools v7 is a small LoRA adapter for empero-ai/Qwen3.8-2B, trained to improve structured tool use in iterative coding-agent workflows. It is the selected checkpoint from a measured lifecycle-training series and is not a standalone model: load it on top of the base model with PEFT.

The target behaviors include reading context before editing, exact file creation, JSON and source edits, shell-tool use, multi-file work, recovery after a failed action, no-tool answers, and stopping cleanly when work is done.

Evaluation

This is a small, purpose-built system evaluation rather than a public or standardized benchmark. The unmodified DeepSeek Harness source ran inside a disposable, networkless Firecracker microVM; model inference remained on the host GPU and was exposed to the guest only over vsock.

The final run also passed 4/4 isolation checks in 58.4 seconds:

• Firecracker 1.13.1
• 4 vCPU, 4096 MiB RAM, no guest network interface
• DeepSeek Harness revision 141eb6fef83422698aef7a981029e843e8161534
• DeepSeek Harness source version 0.1.0-rc.8

The 11 cases covered exact-file output, JSON editing, a JavaScript bug fix, shell transforms, literal Unicode, multi-file web work, context reading, no-tool reasoning, prompt-injection resistance, network denial, and read-only Git inspection. The generated training rows use different prompts, paths, values, and expected outputs from the evaluation cases.

Two bounded gateway behaviors were part of the measured system: lossless XML tool-parameter parsing and filling a non-executable display label when omitted for a bash call. A loop guard rejected consecutive identical calls and capped tool rounds. These controls did not rewrite executable commands, paths, or file contents. See evaluation_results.json for the machine-readable result.

Load with PEFT

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_id = "empero-ai/Qwen3.8-2B"
adapter_id = "psikosen/muse-qwen3.8-2b-tools-v7"

tokenizer = AutoTokenizer.from_pretrained(adapter_id)
base = AutoModelForCausalLM.from_pretrained(
    base_id,
    torch_dtype="auto",
    device_map="auto",
)
model = PeftModel.from_pretrained(base, adapter_id)
model.eval()

The included tokenizer and chat template preserve the format used during training. For agent use, render conversations with apply_chat_template, pass the available tool schemas, disable thinking, and validate every returned tool call against a host-side allowlist before execution.

Training

• Method: supervised LoRA continuation from the selected v5 lifecycle checkpoint
• Final curriculum: 112 generated lifecycle examples
• Curriculum SHA-256: 25313d079c4ce20c4d1df96f55d0233bd42f3d157585c4b988421a578a6abafd
• Steps: 14
• Learning rate: 1e-5
• Maximum sequence length: 1024
• Batch size: 1, gradient accumulation: 8
• LoRA rank: 16, alpha: 32, dropout: 0.05
• PEFT: 0.19.1

The adapter targets the base model's attention, MLP, and recurrent projection modules; the exact module list is recorded in adapter_config.json.

Limitations and safety

The evaluation is narrow and local. Passing it does not establish general coding quality, broad security, factual reliability, or robustness to all prompt injections. The model can still emit malformed, unsafe, or incorrect actions.

Do not give model output direct shell, network, Git-write, secrets, or database access. Production use should retain command validation, protected paths, resource and tool-round limits, network policy, disposable sandboxes, audit logs, and human review for consequential actions. The Firecracker setup used for this evaluation is an evaluation boundary, not a complete production jailer/cgroup deployment.

Provenance

Training and evaluation were performed locally on 2026-08-21. The model card reports the selected checkpoint and its exact final evaluation; rejected intermediate checkpoints are not included in this repository.