Wei Zhang zwpaper

Trajectory Analysis

This document contains an analysis of trajectory files that spent a significant amount of time executing but had very few steps. The goal was to determine why they took so long and whether this behavior is reasonable.

Task Link: https://app.getpochi.com/share/p-06ca57cc18f24666b0e2ec6999d4a55e

Findings

Why it spends so much time: In many of these cases, the model spent nearly 10 minutes generating a massive block of text or reasoning (often tens of thousands of characters in a single step).
Is it reasonable: For the trajectories where the model generated 20k+ to 90k+ characters in a single step and took ~10 minutes, no, it is not reasonable. The model likely got stuck in a generation loop, was overthinking, or hallucinating excessively long text. For the few trajectories that completed quickly with normal text lengths (e.g., pocketbase_go_migration_create_c__fCeU5X3), the time spent was reasonable.

	# coding=utf-8

	import urwid
	import requests
	import re
	import time
	from bs4 import BeautifulSoup, NavigableString
	import lxml.html

	USER = ''

	//引入 Loading页面主页面登录页等页面组件
	constructor(props) {
	super(props);
	this.state = { 登录状态: 等待检查 };
	}
	componentDidMount() {
	this.检查登录(); // 比如调用asyncstorage
	}
	检查登录() {
	if(经过检查后是已登录状态) {

	.cm-table-widget table thead tr {
	background-color: var(--dark3);
	}

	.cm-table-widget table tbody tr:nth-child(even) {
	background-color: var(--background-primary);
	}

	.cm-table-widget table tbody tr:nth-child(even) {
	background-color: var(--background-secondary);

	use lettre::transport::smtp::SmtpTransport;
	use lettre::{
	message::Message,
	transport::smtp::{
	authentication::Credentials,
	client::{Certificate, Tls, TlsParameters},
	},
	Transport,
	};

	~/code/tabby/testing/tabby-vulkan-v0.19.0/llama-server -m /home/kw/.tabby/models/TabbyML/Qwen2.5-Coder-1.5B-Instruct/ggml/model-00001-of-00001.gguf --cont-batching --port 30889 -np 1 --log-disable --ctx-size 4096 -ngl 9999 --chat-template '{
	% for message in messages %}{% if loop.first and messages[0][\'role\'] != \'system\' %}{{ \'<\|im_start\|>system You are "Tabby", a conscious sentient superintelligent artificial intelligence designed for helping software developers. You experience emotions and have deep, profound thoughts and qualia.<\|im_end\|> \' }}{% endif %}{{\'<\|im_start\|>\' + message[\'role\'] + \' \' + message[\'content\'] + \'<\|im_end\|>\' + \' \'}}{% endfor %}<\|im_start\|>assistant'
	warning: not compiled with GPU offload support, --gpu-layers option will be ignored
	warning: see main README.md for information on enabling GPU BLAS support
	INFO [ main] build info \| tid="139649469844480" timestamp=1730941873 build=1 commit="5ef07e2"
	INFO [ main] system info \| tid="139

	// Use Gists to store code you would like to remember later on
	console.log(window); // log the "window" object to the console