AVBala/nimora-agent

Nimora Agent v0.0.1-alpha

Nimora Agent is an experimental QLoRA adapter for Qwen/Qwen3-4B. It is trained to act as the controller in the open-source Nimora coding-agent runtime: select one small, verifiable action at a time, use provider-neutral workspace/Git/change tools, and return canonical JSON decisions.

This release is an alpha protocol adapter, not a general-purpose coding model and not evidence of production-level autonomous software engineering.

Intended use

Use this adapter with the pinned base model and the Nimora runtime. The runtime supplies the system instruction, repository observations, tool execution, policy checks, revision binding, and approval boundaries. Keep a human in control of destructive actions, approvals, merges, and deployments.

The expected response is either:

{"plan":"brief reason","action":{"name":"tool.name","arguments":{}}}

or:

{"result":"evidence-backed final response"}

Evaluation

The publication gate used 40 held-out synthetic prefix-to-next-decision cases across 10 scenario families. The same cases were run with deterministic decoding against the pinned base model and this adapter.

Model	Valid JSON	Correct action/result semantics	Exact action arguments
Qwen3-4B base	100%	27.5%	20.0%
Nimora adapter	100%	100%	97.5%

Final validation loss was 0.004358. All publication gates passed. See evaluation-report.json for every expected decision, raw generation, parsed result, metric, version, and artifact hash.

These results measure conformance to a deterministic synthetic protocol. They do not measure real-repository coding quality and are not comparable to SWE-bench or other independent coding benchmarks.

Training

• Base: Qwen/Qwen3-4B at revision 1cfa9a7208912126459214e8b04321603b3df60c
• Method: 4-bit NF4 QLoRA, rank 16, alpha 32
• Context: 2048 tokens
• Supervision: last assistant turn in balanced prefix-to-next-decision examples
• Data: 400 train, 40 validation, 40 held-out evaluation cases
• Hardware: one Tesla T4
• Train runtime: 805 seconds; total gated run: 1,526 seconds
• License: Apache-2.0

The data is deterministic synthetic template data authored for Nimora. It contains no private repositories, user logs, credentials, or scraped code. Dataset provenance and limitations are recorded in dataset-manifest.json; exact hyperparameters are in training-config.yaml.

Loading

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_id = "Qwen/Qwen3-4B"
adapter_id = "AVBala/nimora-agent"

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

For actual agent inference, use the Nimora runtime rather than sending unconstrained standalone prompts.

Limitations and safety

• The corpus is small and templated; unfamiliar tools, repositories, languages, and long-horizon tasks are unvalidated.
• High held-out template accuracy may not transfer to real coding work.
• The model can still hallucinate, choose unsafe actions, or claim unsupported results outside the trained protocol.
• Never give the model unrestricted shell, credentials, merge, deployment, or approval authority.
• Provider operations should remain revision-bound, policy-checked, auditable, and human-reviewable.

Reproducibility

The repository contains the dataset builder, QLoRA trainer, runtime protocol, provider integrations, and tests. This model repository includes the adapter, tokenizer files, full evaluation report, training configuration, and dataset manifest.