promql-0.5b-onnx

promql-0.5b-onnx

A specialist fine-tune of Qwen2.5-Coder-0.5B that turns natural-language questions into PromQL over a systems-performance metrics schema (CPU, scheduler, block IO, network, syscalls, …), for in-browser inference (transformers.js + WebGPU, ONNX/q4) — a natural-language query box for metrics.

It is schema-agnostic: the prompt supplies candidate metric "cards", and the model answers using only those names — so it generalizes to metrics it never saw in training. Gold PromQL in the training data is execution-validated, never hand-written.

Usage (transformers.js)

import { pipeline } from '@huggingface/transformers';
const gen = await pipeline('text-generation', 'brayniac/promql-0.5b-onnx',
                           { device: 'webgpu', dtype: 'q4' });

Decode greedily and stop on both <|im_end|> and <|endoftext|> (the base eos differs from the ChatML turn terminator). Take the first non-empty line of the completion. Ground the output to the provided card names (snap any out-of-vocab metric identifier to the nearest provided name).

Prompt format (must match exactly)

system

Convert the request into ONE PromQL query using ONLY the listed metrics.
counter -> irate(x[1s]); gauge -> x; histogram -> histogram_quantile(q, x); aggregate with sum()/avg(); filter with {label="value"}; ratio "A per B" -> sum(A_expr)/sum(B_expr).
If no listed metric answers the request, output exactly: NO_METRIC. Output only PromQL, nothing else.

user

Metrics:
  <name> (<type>; labels: <k1>,<k2>) — <description>
  ...                                       # includes 1–3 distractors
Request: <natural language request>

assistant = one PromQL query, or exactly NO_METRIC.

Evaluation (execution-based semantic equivalence)

Intents covered: lookup, total, average, by_label, filter, share (part-of-whole), peak/min, by_index (per-CPU → heatmap), quantile, topk, ratio, derived-efficiency KPIs (IPC/CPI, cache hit/miss rate, branch-miss rate, avg IO/packet size, frequency scaling, cost-per-traffic), and NO_METRIC refusal.

Two things the runtime must uphold

1. Retrieval supplies the cards — for "A per B" / efficiency ratios the retriever MUST include both metrics, or even a perfect model can't answer.
2. Chart type is chosen by the viewer, not the model — from metric_type + result shape (histogram→percentiles/heatmap; sum by (id) multi-series→heatmap; single series→line). The model emits PromQL only.

Files: onnx/model_q4.onnx (int4, shipped) + fp32 model.onnx. A q4f16 (~half the size) is recommended via the transformers.js conversion script.