eventa-1.5b (GGUF)

A small, offline assistant model for Node.js developers. It is a LoRA fine-tune of Qwen2.5-Coder-1.5B-Instruct, quantized to Q4_K_M (about 1 GB). It runs on a laptop CPU.

It is trained for the three tasks of the eventa CLI:

• explain: a Node.js stack trace or TypeScript compile error plus the failing source lines → Cause / Fix / Prevent (Node, Express, NestJS, Prisma)
• review: a git diff plus static-check hints → - [high|medium|low] line N: problem → fix, or No issues found.
• deps: npm audit / outdated / unused / missing facts → a prioritized action plan that keeps the computed fix commands exactly

Use it

With the CLI (recommended). The CLI builds the right prompts for you:

npx @jeeva1398/eventa model use eventa-1.5b
node app.js 2>&1 | npx @jeeva1398/eventa explain
npx @jeeva1398/eventa review
npx @jeeva1398/eventa deps

In GitHub Actions. It reviews every pull request on the runner and posts one comment: uses: Jeeva1398/eventa@v0.2.1. See the repo README.

With Ollama:

ollama run hf.co/jeeva1398/eventa-1.5b-gguf

With llama.cpp:

llama-cli -hf jeeva1398/eventa-1.5b-gguf -sys "You are Eventa, an expert Node.js engineer."

Prompt format

It uses the Qwen ChatML template with this system prompt:

You are Eventa, an expert Node.js engineer. Answer concisely and accurately. Prefer modern Node.js (ESM, async/await, node: built-ins). Use markdown and fenced code blocks.

The user message is a task instruction followed by the context (error output and source, a diff, or dependency facts). For the exact templates, see packages/cli/src/prompts/ in the repo.

Training

• Method: Unsloth QLoRA (r=16, alpha=16, all attention and MLP projections), 2 epochs, lr 2e-4, loss on responses only, run on a free T4 GPU.
• Data: about 740 examples (68 crash types including TypeScript compile errors, NestJS and Prisma; 74 review scenarios: clean diffs, bugs that static checks flag, and bugs only careful reading finds) generated by training/src/build-dataset.ts:
  • real Node.js programs that crash, run and parsed with the CLI's stack-trace parser, with hand-written causes and fixes
  • annotated before/after diffs, including clean diffs where false-positive hints must be dismissed
  • dependency reports built from real advisories and documented major-version breaking changes
• Held out: whole scenarios are kept for evaluation.

Evaluation

54 held-out examples from scenarios never seen in training (npm run eval -w eventa-training). Both models were run through Ollama on CPU with the CLI's prompts.

The base model flags almost every line, so its 100% recall is mostly noise. eventa-1.5b reports only real problems and is about 2× faster because its answers are shorter. Explain quality on unseen error types is slightly below the base model and is the focus of the next data round.

Limitations

• A 1.5B model can still be wrong or vague. Use its answers as a strong hint and verify fixes.
• It is focused on Node.js / JavaScript / TypeScript backend code, and is weaker on browser and framework-specific frontend issues.
• The deps advice relies on facts that the CLI computes. Used without the CLI, it can invent versions.

License

Apache-2.0, the same as the base model Qwen2.5-Coder-1.5B-Instruct (© Alibaba Cloud). The eventa CLI source code is MIT.