Committee

In this article, you will learn seven concrete regression tests for catching the orchestration-layer failure modes that matter most before deploying an AI agent to…

In this tutorial, we implement an end-to-end supervised fine-tuning pipeline for the XYZ-Aquila-SFT dataset, Hugging Face Transformers, PyTorch, and PEFT. We stream and inspect the dataset, parse multi-turn tool-use trajectories, extract structured tool calls, analyze corpus characteristics, and preserve embedded reasoning and observation patterns. We then convert tool schemas between message-embedded and structured formats, render Qwen-compatible ChatML with assistant-only loss masking, prepare a custom PyTorch dataset and collator, and fine-tune Qwen3-0.6B with LoRA. Finally, we evaluate tool-call prediction before and after training and export both the transformed dataset and corpus statistics for further experimentation. Copy CodeCopiedUse a different Browser import os, sys, subprocess CFG = dict( REPO = “XYZAILab/XYZ-Aquila-SFT”, LANG = “en”, N_STREAM = 400, N_EVAL = 40, MODEL_ID = “Qwen/Qwen3-0.6B”, MAX_SEQ_LEN = 2048, LENGTH_POLICY = “truncate”, RUN_TRAINING = True, MAX_STEPS = 30, GRAD_ACCUM = 8, LR = 1e-4, LORA_R = 16, RUN_EVAL = True, N_EVAL_PROBES = 24, OUT_DIR = “/content/aquila_out”, SEED = 0, ) os.makedirs(CFG[“OUT_DIR”], exist_ok=True) def pip(*pkgs): subprocess.run([sys.executable, “-m”, “pip”, “install”, “-q”, “-U”, *pkgs], check=False) pip(“datasets>=3.0.0”, “transformers>=4.51.0”, “peft>=0.13.0”, “accelerate>=1.0.0”) import json, re, math, random, statistics as stats from collections import Counter, defaultdict from dataclasses import dataclass, field from typing import Any, Dict, List, Optional import torch import matplotlib.pyplot as plt from datasets import load_dataset from transformers import AutoTokenizer, AutoModelForCausalLM, get_cosine_schedule_with_warmup random.seed(CFG[“SEED”]); torch.manual_seed(CFG[“SEED”]) DEV = “cuda” if torch.cuda.is_available() else “cpu” BF16 = DEV == “cuda” and torch.cuda.is_bf16_supported() print(f”device={DEV} bf16={BF16} torch={torch.__version__}”) print(f”n[1] streaming {CFG[‘REPO’]}:{CFG[‘LANG’]} …”) stream = load_dataset(CFG[“REPO”], CFG[“LANG”], split=”train”, streaming=True) RAW: List[Dict[str, Any]] = list(stream.take(CFG[“N_STREAM”])) print(f” pulled {len(RAW)} rows; keys = {list(RAW[0].keys())}”) _r = RAW[0] print(f” question[:110] : {_r[‘question’][:110]}…”) print(f” answer : {_r[‘answer’][:80]}”) print(f” number of tool calls : {_r[‘number of tool calls’]}”) print(f” trajectory len : {len(_r[‘trajectory’])} msgs”) print(f” role sequence (first8): {[m[‘role’] for m in _r[‘trajectory’][:8]]}”) We configure the dataset, model, training parameters, output directory, and reproducibility settings for the complete workflow. We install the required Hugging Face, PEFT, Accelerate, and PyTorch-related dependencies and detect whether a CUDA GPU and BF16 support are available. We then stream a limited number of XYZ-Aquila-SFT examples, inspect the dataset schema, and examine the structure of the first tool-use trajectory. Copy CodeCopiedUse a different Browser TOOLS_BLOCK_RE = re.compile(r”<tools>s*(.*?)s*</tools>”, re.S) THINK_RE = re.compile(r”<think>(.*?)</think>”, re.S) TOOL_RESP_RE = re.compile(r”<tool_response>s*(.*?)s*</tool_response>”, re.S) TOOLS_HDR_RE = re.compile(r”nn# Toolsnn”) def iter_json_objects(text: str, limit: int = 1): “””Nesting-safe JSON scanner. Regex like r’\{.*?\}’ breaks on nested `arguments` objects, which every real tool call has.””” dec, i, n, out = json.JSONDecoder(), 0, len(text), [] while i < n and len(out) < limit: while i < n and text[i] not in “{[“: i += 1 if i >= n: break try: obj, end = dec.raw_decode(text, i) except json.JSONDecodeError: i += 1 continue out.append(obj); i = end return out def parse_tool_calls(content: str) -> List[Dict[str, Any]]: calls = [] for m in re.finditer(r”<tool_call>”, content): got = iter_json_objects(content[m.end():], limit=1) if got: calls.append(got[0]) return calls @dataclass class Trajectory: question: str answer: str declared_calls: int messages: List[Dict[str, str]] system_core: str = “” tools: List[Dict[str, Any]] = field(default_factory=list) tools_suffix: str = “” calls: List[Dict[str, Any]] = field(default_factory=list) n_observations: int = 0 n_think: int = 0 @property def tool_names(self): return [c.get(“name”, “?”) for c in self.calls] @property def depth(self): return len(self.messages) def parse_row(row: Dict[str, Any]) -> Trajectory: msgs = [{“role”: m[“role”], “content”: m[“content”]} for m in row[“trajectory”]] t = Trajectory(row[“question”], row[“answer”], row[“number of tool calls”], msgs) if msgs and msgs[0][“role”] == “system”: sysmsg = msgs[0][“content”] split = TOOLS_HDR_RE.search(sysmsg) if split: t.system_core = sysmsg[:split.start()] t.tools_suffix = sysmsg[split.start():] else: t.system_core = sysmsg blk = TOOLS_BLOCK_RE.search(sysmsg) if blk: t.tools = iter_json_objects(blk.group(1), limit=64) for m in msgs: if m[“role”] == “assistant”: t.calls += parse_tool_calls(m[“content”]) t.n_think += len(THINK_RE.findall(m[“content”])) else: t.n_observations += len(TOOL_RESP_RE.findall(m[“content”])) return t TRAJ = [parse_row(r) for r in RAW] t0 = TRAJ[0] print(f”n[2] parsed {len(TRAJ)} trajectories”) print(f” tool schemas found : {[fn.get(‘function’, fn).get(‘name’) for fn in t0.tools]}”) print(f” parsed calls : {len(t0.calls)} (declared {t0.declared_calls})”) print(f” observations : {t0.n_observations} think blocks: {t0.n_think}”) if t0.calls: print(f” sample call : {json.dumps(t0.calls[0], ensure_ascii=False)[:200]}”) agree = sum(len(t.calls) == t.declared_calls for t in TRAJ) print(f” parser vs ‘number of tool calls’: {agree}/{len(TRAJ)} exact match”) calls_per = [len(t.calls) for t in TRAJ] depth_per = [t.depth for t in TRAJ] chars_per = [sum(len(m[“content”]) for m in t.messages) for t in TRAJ] name_freq = Counter(n for t in TRAJ for n in t.tool_names) argkey_freq = defaultdict(Counter) for t in TRAJ: for c in t.calls: args = c.get(“arguments”, {}) if isinstance(args, dict): for k in args: argkey_freq[c.get(“name”, “?”)][k] += 1 def q(xs, p): xs = sorted(xs); return xs[min(len(xs) – 1, int(p * len(xs)))] print(“n[3] corpus statistics”) print(f” tool calls / traj : mean {stats.mean(calls_per):.1f} p50 {q(calls_per,.5)} ” f”p90 {q(calls_per,.9)} max {max(calls_per)}”) print(f” messages / traj : mean {stats.mean(depth_per):.1f} p90 {q(depth_per,.9)} max {max(depth_per)}”) print(f” chars / traj : mean {stats.mean(chars_per):,.0f} p90 {q(chars_per,.9):,}”) print(f” tool distribution : {dict(name_freq)}”) for k, v in argkey_freq.items(): print(f” {k:<24} arg keys -> {dict(v.most_common(6))}”) tot = sum(chars_per); top = sum(sorted(chars_per)[-max(1, len(chars_per)//10):]) print(f” top-10% longest trajectories hold {100*top/tot:.1f}% of all characters”) fig, ax = plt.subplots(1, 3, figsize=(15, 3.6)) ax[0].hist(calls_per, bins=40); ax[0].set_yscale(“log”); ax[0].set_title(“tool calls / trajectory”) ax[1].hist(depth_per, bins=40); ax[1].set_yscale(“log”); ax[1].set_title(“messages / trajectory”) ax[2].bar(list(name_freq), list(name_freq.values())); ax[2].set_title(“tool usage”); ax[2].tick_params(axis=”x”, rotation=20) plt.tight_layout(); plt.show() We define nesting-safe utilities for extracting JSON tool calls, reasoning blocks, observations, and embedded tool schemas from each conversation. We convert every raw dataset row into a structured trajectory object and verify that the parsed tool-call counts match the values declared by the dataset. We then calculate corpus-level statistics and visualize the distributions of tool calls, message depth, trajectory size, and tool usage frequency. Copy CodeCopiedUse a different Browser QWEN3_TOOLS_TMPL = ( “You are provided with function signatures within <tools></tools> XML tags:n<tools>n” “{lines}n</tools>nnFor each function call, return a json object with function name ” “and arguments within <tool_call></tool_call> XML tags:n<tool_call>n” ‘{{“name”: <function-name>, “arguments”: <args-json-object>}}n</tool_call>’ ) def extract_tools(t: Trajectory) -> Dict[str, Any]: “””message-embedded schemas -> {‘messages’: […], ‘tools’: […]}””” msgs = [dict(m) for m in t.messages] if msgs and msgs[0][“role”] == “system”: msgs[0][“content”] = t.system_core return {“messages”: msgs, “tools”: t.tools, “question”: t.question, “answer”: t.answer} def render_tools(rec: Dict[str, Any]) -> List[Dict[str, str]]: “””inverse: structured tools -> schemas re-embedded in the system message”””

In 2017, Deanne Taylor attended a presentation at the University of Pennsylvania, just a short walk from her office. A researcher was there to unveil the Human Cell Atlas, an ambitious project that aimed to map every cell in the human body. Taylor was floored, and then concerned. As details emerged, she discovered that the project’s researchers had only made plans to study adults. “That’s when my little alarm went off,” she says. “Not again.” Since joining the Children’s Hospital of Philadelphia (CHOP) as the director of bioinformatics three years earlier, Taylor had been disappointed by the lack of investment in medical research focused on children. The dominant view, she says, was that children are exactly like small adults. They’re not. Children’s cells are different from grownups’ cells in the way they express genes—switching them on and off or turning them up or down. Those variations can cause drastically different and even deadly responses to drugs that adults tolerate well.  The 2017 talk was the moment Taylor didn’t know she’d been waiting for. She quickly channeled her concern into a campaign, joining the Human Cell Atlas’s volunteer team and helping write a section on children for a white paper outlining the group’s goals and plans. She then rallied a cross-hospital coalition of pediatric researchers to contribute to the project and spearheaded a 2019 paper that outlined the case for studying children—a bid to attract more interest and funding to the field. “It put a flag in the ground,” she says. “Why don’t we have healthy models of children’s development?” So far, the push has paid off. In 2021 the NIH awarded a $38.5 million grant to the Developmental Genotype-Tissue Expression Project (dGTEx), a major initiative aimed at establishing the first comprehensive database of healthy pediatric tissue. The project banks samples collected from otherwise healthy children who have died and whose parents agreed to donate their bodies, and maps how genes across all the major organ systems are expressed. Taylor and her team curate and standardize the information associated with each tissue donation, including family history and details about the samples. A separate group does analysis on the samples themselves, and then all the information is combined to create a database­—a baseline of what gene expression looks like in children. It’s the first step to enabling research that could advance our knowledge of normal development, disease, drug effectiveness, and other phenomena.  The dGTEx team will eventually feed its data into the Human Cell Atlas, which, thanks to Taylor and many of the coauthors of the 2019 paper, now includes a pediatric section. Taylor’s primary responsibility may be collecting and organizing data for dGTEx, but colleagues say she’s also the glue holding diverse research projects together. That’s especially important for the Human Cell Atlas, which depends on contributions from a loose coalition of researchers, all pursuing their own objectives. “Deanne took a big-picture view and said, We don’t just need to understand the pediatric kidney or the pediatric brain or the pediatric immune system. We need a holistic view of pediatric development,” says Sarah Teichmann, a cofounder of the Human Cell Atlas. “She embodies that interdisciplinary spirit.”  A healthy baseline Taylor describes her career as a “random walk,” driven by a singular intensity she now attributes to undiagnosed autism and ADHD. At five, she began reading her mom’s medical texts. By 12, she was checking out physics books from the library. Physics provided mysteries to solve, and she wanted to understand how things worked.  Taylor got her PhD in biophysics, in 2001, but was inspired by the then-active Human Genome Project to change gears and take on a postdoc at Pfizer, writing code to handle complex data in rare-disease research. Then she moved to reproductive medicine, where she worked on some of the first computer programs to screen embryos for chromosomal abnormalities—many of which are still in use today. Despite this seemingly winding road, Taylor says her focus has always been on understanding why the same illness hits people differently. How can two people carry the same disease-­associated gene variant, but only one get sick?  The Human Cell Atlas—including all the data feeding into it from dGTEx and other projects—could at last help researchers find answers. The effort is a natural extension of the Human Genome Project. That initiative, which wrapped up in 2003, helped researchers link specific genes to specific diseases. But a map of the genome is a bit like a DIY kit with all the parts and no assembly manual. It doesn’t tell you where and how cells use each gene throughout the body. After all, “we’re just older kids,” Taylor says. “By ignoring the pediatric side of things, I think people are missing a window of intervention in human disease.” For that, you need to know how the genes are expressed. Gene expression generally involves making a protein that does a specific job in the body, like building tissue or sending signals. Unlike DNA, which largely remains the same throughout our lives, the way the genes in DNA are expressed changes as we develop.  Differences in gene expression can determine whether a therapy will work—or could harm more than it helps. Because of the way cardiac genes are expressed in children, chemotherapy drugs can attack not only tumors but also children’s developing hearts, potentially causing lifelong damage. Other treatments can affect the entire body, sometimes triggering a reversible but potentially fatal immune-system reaction called cytokine release syndrome. The dGTEx database aims to create a baseline for gene expression in children—a molecular map of how the body’s roughly 20,000 genes do their work in healthy tissue cells. It is only one of the collaborations Taylor manages. She’s a principal investigator for the Kids First Data Resource Center, which sequences diseased tissues collected from children enrolled in other studies nationwide. And she has been collaborating with researchers on HubMAP, an effort that’s building a resource complementary to the Human Cell Atlas, to secure funding to create 3D maps of

This week I spoke to scientists who have found a way to turn male mouse embryos female. They’ve developed a CRISPR-based approach to essentially cut out the Y chromosome. It allowed them to create female clones of male mice. That’s right: female animals that are genetically identical to males, except for the missing Y chromosome. Takashi Ishiuchi, a reproductive biologist at the University of Yamanashi who co-led the work, told me it felt a bit like sci-fi. Ishiuchi and his colleague Shogo Matoba of the Riken BioResource Research Center hope their approach could be helpful in conservation efforts, especially in cases where we might have only a few individuals of a species left. But cloning has multiple uses, ranging from the cool to the outright creepy. We can’t talk about cloning without mentioning Dolly, the celebrity sheep born in 1996 and the first mammal successfully cloned from an adult cell. In that case, scientists took the DNA-containing nucleus of an adult mammary cell from one sheep and transferred it into an egg cell that had had its own nucleus removed. The resulting embryo was transferred to a surrogate sheep, which gave birth to Dolly—an animal genetically identical to the DNA donor. The scientists behind that work were interested in genetically modifying livestock. Farmers have essentially been doing this for thousands of years through selective breeding, but cloning allows scientists to create genetic replicas of animals with desirable traits. Cloning is also being used to replicate deceased pets, including, famously, those of Barbra Streisand and Tom Brady, among others. For a price somewhere in the tens of thousands of dollars, a company can take cells from your pet and turn them into a living, breathing clone. Considering that cloning also requires egg cells from another animal, and a surrogate animal to carry the pregnancy, not everyone is on board with this, especially since there is no medical or environmental need for the procedures. One bioethicist, Jessica Pierce, has described this aspect of dog cloning as “the exploitation of the canine underclass.” The case for cloning is stronger when it comes to conservation—where some argue there is environmental value. Scientists have been preserving animal tissues for years. Some of these tissues are cryopreserved at low temperatures in “frozen zoos.” The facility at the San Diego Zoo, for example, currently has cells from over 1,300 species. Some of these samples were taken decades ago. Preserved tissues like these have enabled scientists to create clones of animals considered close to extinction, including black-footed ferrets and Przewalski’s horse. But they might also help us bring back extinct animals. In 2009, researchers in Spain described how they’d cloned an extinct wild goat, the Pyrenean ibex, using skin cells that had been cryopreserved a decade earlier. In that research, the team used egg cells from domestic goats to create a total of 439 embryos. Ultimately, only one goat—a female—was born. She died minutes later because of a defect in her lungs. Poor Pyrenean ibex. It’s the only animal we know of that has gone extinct twice. The biotech company Colossal Biosciences is hoping to use old—and potentially ancient—genetic material to bring back long-extinct species like the thylacine and woolly mammoth. So far, the company’s efforts have largely involved modifying the genomes of modern-day animals. Technically, it’s also possible to clone humans. As far as we know, no one has done it. But some have played with the idea. One biotech startup founder has pitched an idea for “brainless clones”—human clones that lack a brain but contain all the organs people might need to replace their own in future. My colleague Antonio Regalado described that pitch in March. (I had to pause eating my lunch while rereading it.) Scientists have done a hell of a lot with cloning over the last few decades. I’m excited—but also slightly nervous—about what the coming decades will bring. This article first appeared in The Checkup, MIT Technology Review’s weekly biotech newsletter. To receive it in your inbox every Thursday, and read articles like this first, sign up here.

In this article, you will learn what latent spaces are and how they serve three distinct roles — descriptive, generative, and predictive — across a…

Listen to the session or watch below The “censorship-industrial complex” is an idea that a network of government, tech, and research groups is collaborating to suppress conservative online speech. This was fodder for the right-wing information sphere for years—then it began making its way into US policy. Watch a conversation exploring how it started, where it’s going, and what it means for the future of democracy and the internet. Speakers: Amy Nordrum, Executive Editor, Operations, and Eileen Guo, Senior Reporter, Features & Investigations Recorded on August 13, 2026 Related Story: How ideas of a vast censorship network moved from the online fringe to Trump policy