精选80MarkTechPost（RSS）技巧与观点
从推理语料到推理模型：流式处理、筛选与微调指南
Create a Reasoning-Focused LLM: A Practical Guide to Streaming, Curating, and Fine-Tuning the SupraLabs Reasoning Corpus
推荐理由
做推理模型微调的同学必看，这份指南把数据流式采样、质量过滤到 LoRA 微调的完整流程都跑通了，还附了可直接照做的代码，赶紧拿去复现。
In this tutorial, we build an end-to-end workflow for working with the SupraLabs reasoning corpus. We stream a representative subset directly from the Hugging Face Hub, inspect its source distribution, token-length patterns, task composition, and reasoning-to-answer ratios, and then apply a series of quality filters to remove unsuitable training examples. We transform the retained samples into a chat-based supervised fine-tuning format with explicit <think> reasoning tags and use them to adapt SmolLM2-135M-Instruct with LoRA through TRL’s SFTTrainer. By combining scalable data access, exploratory analysis, dataset curation, parameter-efficient fine-tuning, structured inference, and Parquet export, we create a complete Google Colab pipeline for turning a large multi-model reasoning corpus into a compact reasoning-focused language model.
在本教程中，我们构建了一个端到端的工作流程来处理SupraLabs推理语料库。我们直接从Hugging Face Hub流式传输一个具有代表性的子集，检查其来源分布、令牌长度模式、任务组成和推理与答案的比例，然后应用一系列质量过滤器来移除不合适的训练示例。我们将保留的样本转换为带有显式<think>推理标签的基于聊天的监督微调格式，并通过TRL的SFTTrainer使用LoRA来适配SmolLM2-135M-Instruct。通过结合可扩展的数据访问、探索性分析、数据集整理、参数高效微调、结构化推理和Parquet导出，我们创建了一个完整的Google Colab流水线，用于将大型多模型推理语料库转化为紧凑的、专注于推理的语言模型。
代码 · 45 行
import subprocess, sys
def pip_install(pkgs):
subprocess.check_call([sys.executable, "-m", "pip", "install", "-q", *pkgs])
subprocess.call([sys.executable, "-m", "pip", "uninstall", "-y", "-q", "torchao"])
pip_install([
"datasets>=3.0.0",
"transformers>=4.46.0",
"trl>=0.12.0",
"peft>=0.13.0",
"accelerate>=1.0.0",
"bitsandbytes",
"matplotlib",
"pandas",
])
import os, re, json, math, random, itertools, warnings
import pandas as pd
import matplotlib.pyplot as plt
import torch
from collections import Counter
from datasets import load_dataset, Dataset
warnings.filterwarnings("ignore")
random.seed(42)
torch.manual_seed(42)
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
print(f"Device: {DEVICE}")
if DEVICE == "cuda":
print(f"GPU: {torch.cuda.get_device_name(0)}")
DATASET_ID = "SupraLabs/reasoning-corpus-4K-5M-v1"
SAMPLE_SIZE = 8_000
print(f"\nStreaming {DATASET_ID} ...")
stream = load_dataset(DATASET_ID, split="train", streaming=True)
stream = stream.shuffle(seed=42, buffer_size=30_000)
rows = list(itertools.islice(stream, SAMPLE_SIZE))
ds = Dataset.from_list(rows)
print(f"Materialized sample: {len(ds):,} rows")
print(f"Columns: {ds.column_names}")
ex = ds[0]
print("\n" + "=" * 70)
print("EXAMPLE ROW")
print("=" * 70)
print(f"repo_id : {ex['repo_id']}")
print(f"tok_len : {ex['tok_len']}")
print(f"user : {ex['user'][:300]} ...")
print(f"thought_trace : {ex['thought_trace'][:300]} ...")
print(f"assistant : {ex['assistant'][:300]} ...")代码 · 45 行
import subprocess, sys
def pip_install(pkgs):
subprocess.check_call([sys.executable, "-m", "pip", "install", "-q", *pkgs])
subprocess.call([sys.executable, "-m", "pip", "uninstall", "-y", "-q", "torchao"])
pip_install([
"datasets>=3.0.0",
"transformers>=4.46.0",
"trl>=0.12.0",
"peft>=0.13.0",
"accelerate>=1.0.0",
"bitsandbytes",
"matplotlib",
"pandas",
])
import os, re, json, math, random, itertools, warnings
import pandas as pd
import matplotlib.pyplot as plt
import torch
from collections import Counter
from datasets import load_dataset, Dataset
warnings.filterwarnings("ignore")
random.seed(42)
torch.manual_seed(42)
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
print(f"Device: {DEVICE}")
if DEVICE == "cuda":
print(f"GPU: {torch.cuda.get_device_name(0)}")
DATASET_ID = "SupraLabs/reasoning-corpus-4K-5M-v1"
SAMPLE_SIZE = 8_000
print(f"\nStreaming {DATASET_ID} ...")
stream = load_dataset(DATASET_ID, split="train", streaming=True)
stream = stream.shuffle(seed=42, buffer_size=30_000)
rows = list(itertools.islice(stream, SAMPLE_SIZE))
ds = Dataset.from_list(rows)
print(f"Materialized sample: {len(ds):,} rows")
print(f"Columns: {ds.column_names}")
ex = ds[0]
print("\n" + "=" * 70)
print("EXAMPLE ROW")
print("=" * 70)
print(f"repo_id : {ex['repo_id']}")
print(f"tok_len : {ex['tok_len']}")
print(f"user : {ex['user'][:300]} ...")
print(f"thought_trace : {ex['thought_trace'][:300]} ...")
print(f"assistant : {ex['assistant'][:300]} ...")We configure the Colab environment, install the required machine learning libraries, and remove the incompatible torchao package. We detect the available compute device, connect to the SupraLabs reasoning corpus through Hugging Face streaming, and avoid downloading the complete dataset. We shuffle the streamed records, materialize a representative sample, and inspect the structure and contents of an example row.
我们配置Colab环境，安装所需的机器学习库，并移除不兼容的torchao包。我们检测可用的计算设备，通过Hugging Face流式传输连接到SupraLabs推理语料库，并避免下载完整数据集。我们打乱流式记录，物化一个代表性样本，并检查示例行的结构和内容。
代码 · 40 行
df = ds.to_pandas()
print("\nTop 15 source repos in sample:")
src_counts = df["repo_id"].value_counts()
print(src_counts.head(15).to_string())
fig, axes = plt.subplots(2, 2, figsize=(14, 10))
axes[0, 0].hist(df["tok_len"], bins=60, color="#4C72B0", edgecolor="white")
axes[0, 0].set_title("Token length distribution")
axes[0, 0].set_xlabel("tok_len"); axes[0, 0].set_ylabel("rows")
src_counts.head(12).plot(kind="barh", ax=axes[0, 1], color="#55A868")
axes[0, 1].invert_yaxis()
axes[0, 1].set_title("Top-12 source repos (sample)")
df["think_chars"] = df["thought_trace"].str.len()
df["answer_chars"] = df["assistant"].str.len()
df["reason_ratio"] = df["think_chars"] / (df["think_chars"] + df["answer_chars"] + 1)
axes[1, 0].hist(df["reason_ratio"], bins=50, color="#C44E52", edgecolor="white")
axes[1, 0].set_title("Reasoning ratio (think / (think + answer))")
axes[1, 0].set_xlabel("ratio")
axes[1, 1].scatter(df["tok_len"], df["reason_ratio"], s=4, alpha=0.25, color="#8172B2")
axes[1, 1].set_title("tok_len vs reasoning ratio")
axes[1, 1].set_xlabel("tok_len"); axes[1, 1].set_ylabel("ratio")
plt.tight_layout()
plt.show()
print("\nSummary stats:")
print(df[["tok_len", "think_chars", "answer_chars", "reason_ratio"]]
.describe().round(2).to_string())
def tag_task(row):
u = row["user"].lower()
a = row["assistant"]
if "```" in a or re.search(r"\b(def |class |import |function|#include)", a):
return "code"
if re.search(r"(prove|equation|integral|theorem|\\frac|\\int|solve for)", u):
return "math"
if re.search(r"\b(patient|diagnosis|symptom|treatment|clinical)\b", u):
return "medical"
if re.search(r"\b(which of the following|options?:|\(a\)|\(b\))", u):
return "mcq/logic"
return "general"
df["task"] = df.apply(tag_task, axis=1)
print("\nHeuristic task mix:")
print(df["task"].value_counts(normalize=True).round(3).to_string())代码 · 40 行
df = ds.to_pandas()
print("\nTop 15 source repos in sample:")
src_counts = df["repo_id"].value_counts()
print(src_counts.head(15).to_string())
fig, axes = plt.subplots(2, 2, figsize=(14, 10))
axes[0, 0].hist(df["tok_len"], bins=60, color="#4C72B0", edgecolor="white")
axes[0, 0].set_title("Token length distribution")
axes[0, 0].set_xlabel("tok_len"); axes[0, 0].set_ylabel("rows")
src_counts.head(12).plot(kind="barh", ax=axes[0, 1], color="#55A868")
axes[0, 1].invert_yaxis()
axes[0, 1].set_title("Top-12 source repos (sample)")
df["think_chars"] = df["thought_trace"].str.len()
df["answer_chars"] = df["assistant"].str.len()
df["reason_ratio"] = df["think_chars"] / (df["think_chars"] + df["answer_chars"] + 1)
axes[1, 0].hist(df["reason_ratio"], bins=50, color="#C44E52", edgecolor="white")
axes[1, 0].set_title("Reasoning ratio (think / (think + answer))")
axes[1, 0].set_xlabel("ratio")
axes[1, 1].scatter(df["tok_len"], df["reason_ratio"], s=4, alpha=0.25, color="#8172B2")
axes[1, 1].set_title("tok_len vs reasoning ratio")
axes[1, 1].set_xlabel("tok_len"); axes[1, 1].set_ylabel("ratio")
plt.tight_layout()
plt.show()
print("\nSummary stats:")
print(df[["tok_len", "think_chars", "answer_chars", "reason_ratio"]]
.describe().round(2).to_string())
def tag_task(row):
u = row["user"].lower()
a = row["assistant"]
if "```" in a or re.search(r"\b(def |class |import |function|#include)", a):
return "code"
if re.search(r"(prove|equation|integral|theorem|\\frac|\\int|solve for)", u):
return "math"
if re.search(r"\b(patient|diagnosis|symptom|treatment|clinical)\b", u):
return "medical"
if re.search(r"\b(which of the following|options?:|\(a\)|\(b\))", u):
return "mcq/logic"
return "general"
df["task"] = df.apply(tag_task, axis=1)
print("\nHeuristic task mix:")
print(df["task"].value_counts(normalize=True).round(3).to_string())We convert the sampled dataset into a pandas DataFrame and analyze the distribution of source repositories and token lengths. We calculate reasoning and answer character counts, measure the reasoning-to-response ratio, and visualize the relationships across the dataset. We also apply lightweight heuristic rules to classify each record as a code, mathematics, medical, multiple-choice, or general task.
我们将采样数据集转换为pandas DataFrame，并分析源仓库和令牌长度的分布。我们计算推理和答案的字符数，衡量推理与响应的比率，并可视化数据集中的关系。我们还应用轻量级启发式规则，将每条记录分类为代码、数学、医学、多项选择或一般任务。
代码 · 51 行
def filter_length(row, min_tok=200, max_tok=3000):
"""Keep samples within a training-friendly token budget."""
return min_tok <= row["tok_len"] <= max_tok
def filter_degenerate(row):
"""Drop empty/near-empty thoughts or answers."""
return len(row["thought_trace"]) > 100 and len(row["assistant"]) > 20
def filter_repetition(row, max_line_repeat=0.30):
"""Drop traces where one line repeats too often (looping models)."""
lines = [l.strip() for l in row["thought_trace"].split("\n") if l.strip()]
if len(lines) < 5:
return True
most_common = Counter(lines).most_common(1)[0][1]
return (most_common / len(lines)) <= max_line_repeat
def filter_reason_ratio(row, lo=0.15, hi=0.97):
"""Keep samples that actually reason but don't ONLY reason."""
t, a = len(row["thought_trace"]), len(row["assistant"])
r = t / (t + a + 1)
return lo <= r <= hi
n0 = len(ds)
ds_f = ds.filter(filter_length)
ds_f = ds_f.filter(filter_degenerate)
ds_f = ds_f.filter(filter_repetition)
ds_f = ds_f.filter(filter_reason_ratio)
print(f"\nFiltering: {n0:,} -> {len(ds_f):,} rows "
f"({100 * len(ds_f) / n0:.1f}% retained)")
MODEL_ID = "HuggingFaceTB/SmolLM2-135M-Instruct"
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
if tokenizer.pad_token is None:
tokenizer.pad_token = tokenizer.eos_token
SYSTEM_PROMPT = (
"You are a careful reasoning assistant. Think step by step inside "
"<think>...</think> tags, then give your final answer."
)
def to_chat(row):
return {
"messages": [
{"role": "system", "content": SYSTEM_PROMPT},
{"role": "user", "content": row["user"]},
{"role": "assistant",
"content": f"<think>\n{row['thought_trace']}\n</think>\n\n{row['assistant']}"},
]
}
train_ds = ds_f.map(to_chat, remove_columns=ds_f.column_names)
train_ds = train_ds.shuffle(seed=42)
N_TRAIN, N_EVAL = 1_500, 100
eval_ds = train_ds.select(range(N_TRAIN, min(N_TRAIN + N_EVAL, len(train_ds))))
train_ds = train_ds.select(range(min(N_TRAIN, len(train_ds))))
print(f"\nTrain: {len(train_ds):,} | Eval: {len(eval_ds):,}")
print("\nRendered training sample (truncated):")
print(tokenizer.apply_chat_template(train_ds[0]["messages"], tokenize=False)[:800])代码 · 51 行
def filter_length(row, min_tok=200, max_tok=3000):
"""Keep samples within a training-friendly token budget."""
return min_tok <= row["tok_len"] <= max_tok
def filter_degenerate(row):
"""Drop empty/near-empty thoughts or answers."""
return len(row["thought_trace"]) > 100 and len(row["assistant"]) > 20
def filter_repetition(row, max_line_repeat=0.30):
"""Drop traces where one line repeats too often (looping models)."""
lines = [l.strip() for l in row["thought_trace"].split("\n") if l.strip()]
if len(lines) < 5:
return True
most_common = Counter(lines).most_common(1)[0][1]
return (most_common / len(lines)) <= max_line_repeat
def filter_reason_ratio(row, lo=0.15, hi=0.97):
"""Keep samples that actually reason but don't ONLY reason."""
t, a = len(row["thought_trace"]), len(row["assistant"])
r = t / (t + a + 1)
return lo <= r <= hi
n0 = len(ds)
ds_f = ds.filter(filter_length)
ds_f = ds_f.filter(filter_degenerate)
ds_f = ds_f.filter(filter_repetition)
ds_f = ds_f.filter(filter_reason_ratio)
print(f"\nFiltering: {n0:,} -> {len(ds_f):,} rows "
f"({100 * len(ds_f) / n0:.1f}% retained)")
MODEL_ID = "HuggingFaceTB/SmolLM2-135M-Instruct"
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
if tokenizer.pad_token is None:
tokenizer.pad_token = tokenizer.eos_token
SYSTEM_PROMPT = (
"You are a careful reasoning assistant. Think step by step inside "
"<think>...</think> tags, then give your final answer."
)
def to_chat(row):
return {
"messages": [
{"role": "system", "content": SYSTEM_PROMPT},
{"role": "user", "content": row["user"]},
{"role": "assistant",
"content": f"<think>\n{row['thought_trace']}\n</think>\n\n{row['assistant']}"},
]
}
train_ds = ds_f.map(to_chat, remove_columns=ds_f.column_names)
train_ds = train_ds.shuffle(seed=42)
N_TRAIN, N_EVAL = 1_500, 100
eval_ds = train_ds.select(range(N_TRAIN, min(N_TRAIN + N_EVAL, len(train_ds))))
train_ds = train_ds.select(range(min(N_TRAIN, len(train_ds))))
print(f"\nTrain: {len(train_ds):,} | Eval: {len(eval_ds):,}")
print("\nRendered training sample (truncated):")
print(tokenizer.apply_chat_template(train_ds[0]["messages"], tokenize=False)[:800])更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力