Spaces:

Devseis
/

Arx

Running

Arx / mobile /test /webllm-test.html

Security assistant uses the fine-tuned Devseis model v0.4 (1.5B)

53e9e4a verified 9 days ago

2.66 kB

<!doctype html><meta charset="utf-8"><title>WebLLM model test</title>
<script type="module">
// Loads a Devseis auditor model build in WebLLM and answers prompts given by mobile/test/webllm-test.cjs.
import * as webllm from '../../app/node_modules/@mlc-ai/web-llm/lib/index.js';
import { SYSTEM_PROMPT, groundingProblems, ANSWER_SCHEMAS } from '../../app/renderer/auditor-core.js';
import { chatProblems } from '../../app/renderer/knowledge.js';
window.loadModel = async (modelUrl) => {
  const adapter = await navigator.gpu.requestAdapter();
  const f16 = adapter.features.has('shader-f16');
  const named = (modelUrl.match(/q0f16|q4f16_1|q4f32_1/) || [])[0];
  const quant = named || (f16 ? 'q4f16_1' : 'q4f32_1');
  const size = /1\.5b/i.test(modelUrl) ? '1.5B' : '0.5B';
  const base = webllm.prebuiltAppConfig.model_list.find(m => m.model_id === `Qwen2.5-${size}-Instruct-${quant}-MLC`);
  const url = modelUrl.replace('{quant}', quant);
  const custom = { ...base, model: url, model_id: 'devseis-test-' + quant };
  // ?ctx=N caps the KV cache (context window); the 1.5B default (32768) breaks after the first answer on some GPUs.
  const ctx = +new URLSearchParams(location.search).get('ctx') || 0;
  window.engine = await webllm.CreateMLCEngine(custom.model_id, { appConfig: { ...webllm.prebuiltAppConfig, model_list: [custom] } },
    ctx ? { context_window_size: ctx } : undefined);
  return { quant, url };
};
// task 'chat': messages = the full chat messages (system + user); otherwise prompt = report prompt.
window.ask = async (task, prompt, useSchema) => {
  if (new URLSearchParams(location.search).get('reset')) await window.engine.resetChat();
  if (task === 'chat') {
    const t0 = performance.now();
    const r = await window.engine.chat.completions.create({ messages: prompt, temperature: 0, max_tokens: 700 });
    const reply = r.choices[0].message.content;
    return { seconds: Math.round((performance.now() - t0) / 1000), problems: chatProblems(prompt[prompt.length - 1].content, reply), reply };
  }
  const t = performance.now();
  const r = await window.engine.chat.completions.create({ messages: [{ role: 'system', content: SYSTEM_PROMPT }, { role: 'user', content: prompt }],
    temperature: 0, max_tokens: 900, ...(useSchema ? { response_format: { type: 'json_object', schema: JSON.stringify(ANSWER_SCHEMAS[task]) } } : {}) });
  const reply = r.choices[0].message.content;
  let problems;
  try { problems = groundingProblems(task, prompt, JSON.parse(reply)); } catch (e) { problems = ['invalid JSON']; }
  return { seconds: Math.round((performance.now() - t) / 1000), problems, reply };
};
window.ready = true;
</script>

	<!doctype html><meta charset="utf-8"><title>WebLLM model test</title>
	<script type="module">
	// Loads a Devseis auditor model build in WebLLM and answers prompts given by mobile/test/webllm-test.cjs.
	import * as webllm from '../../app/node_modules/@mlc-ai/web-llm/lib/index.js';
	import { SYSTEM_PROMPT, groundingProblems, ANSWER_SCHEMAS } from '../../app/renderer/auditor-core.js';
	import { chatProblems } from '../../app/renderer/knowledge.js';
	window.loadModel = async (modelUrl) => {
	const adapter = await navigator.gpu.requestAdapter();
	const f16 = adapter.features.has('shader-f16');
	const named = (modelUrl.match(/q0f16\|q4f16_1\|q4f32_1/) \|\| [])[0];
	const quant = named \|\| (f16 ? 'q4f16_1' : 'q4f32_1');
	const size = /1\.5b/i.test(modelUrl) ? '1.5B' : '0.5B';
	const base = webllm.prebuiltAppConfig.model_list.find(m => m.model_id === `Qwen2.5-${size}-Instruct-${quant}-MLC`);
	const url = modelUrl.replace('{quant}', quant);
	const custom = { ...base, model: url, model_id: 'devseis-test-' + quant };
	// ?ctx=N caps the KV cache (context window); the 1.5B default (32768) breaks after the first answer on some GPUs.
	const ctx = +new URLSearchParams(location.search).get('ctx') \|\| 0;
	window.engine = await webllm.CreateMLCEngine(custom.model_id, { appConfig: { ...webllm.prebuiltAppConfig, model_list: [custom] } },
	ctx ? { context_window_size: ctx } : undefined);
	return { quant, url };
	};
	// task 'chat': messages = the full chat messages (system + user); otherwise prompt = report prompt.
	window.ask = async (task, prompt, useSchema) => {
	if (new URLSearchParams(location.search).get('reset')) await window.engine.resetChat();
	if (task === 'chat') {
	const t0 = performance.now();
	const r = await window.engine.chat.completions.create({ messages: prompt, temperature: 0, max_tokens: 700 });
	const reply = r.choices[0].message.content;
	return { seconds: Math.round((performance.now() - t0) / 1000), problems: chatProblems(prompt[prompt.length - 1].content, reply), reply };
	}
	const t = performance.now();
	const r = await window.engine.chat.completions.create({ messages: [{ role: 'system', content: SYSTEM_PROMPT }, { role: 'user', content: prompt }],
	temperature: 0, max_tokens: 900, ...(useSchema ? { response_format: { type: 'json_object', schema: JSON.stringify(ANSWER_SCHEMAS[task]) } } : {}) });
	const reply = r.choices[0].message.content;
	let problems;
	try { problems = groundingProblems(task, prompt, JSON.parse(reply)); } catch (e) { problems = ['invalid JSON']; }
	return { seconds: Math.round((performance.now() - t) / 1000), problems, reply };
	};
	window.ready = true;
	</script>