rprint(Panel.fit("[bold]Baseline 1: Predict output_type from context using pure Python Naive Bayes[/bold]"))
model_artifacts = {}
classifier_df = df.dropna(subset=["output_type"]).copy()
classifier_df = classifier_df[
classifier_df["output_type"].astype(str).str.len() > 0
].copy()
if classifier_df["output_type"].nunique() >= 2 and len(classifier_df) >= 30:
X_text = (
classifier_df["context"]
.fillna("")
.astype(str)
.map(lambda text: text[:12000])
.tolist()
)
y = classifier_df["output_type"].astype(str).tolist()
train_indices, test_indices = stratified_train_test_indices(y, test_size=0.2, seed=SEED)
X_train = [X_text[i] for i in train_indices]
y_train = [y[i] for i in train_indices]
X_test = [X_text[i] for i in test_indices]
y_test = [y[i] for i in test_indices]
output_type_classifier = PureMultinomialNB(
max_features=20000,
min_df=2,
alpha=1.0,
)
output_type_classifier.fit(X_train, y_train)
predictions = output_type_classifier.predict(X_test)
output_type_metrics, output_report_df = evaluate_predictions(y_test, predictions)
output_matrix_df = confusion_matrix_df(y_test, predictions)
output_type_metrics["train_rows"] = len(X_train)
output_type_metrics["test_rows"] = len(X_test)
output_type_metrics["vocab_size"] = len(output_type_classifier.vocab)
rprint("[bold]Output type classifier report:[/bold]")
display(output_report_df)
display(output_matrix_df)
output_report_df.to_csv(OUT_DIR / "output_type_classifier_report.csv", index=False)
output_matrix_df.to_csv(OUT_DIR / "output_type_confusion_matrix.csv")
top_token_records = []
for label in output_type_classifier.labels:
for token, margin in output_type_classifier.top_tokens_for_class(label, n=25):
top_token_records.append(
{
"label": label,
"token": token,
"score_margin": margin,
}
)
pd.DataFrame(top_token_records).to_csv(
OUT_DIR / "output_type_top_tokens.csv",
index=False,
)
with open(
OUT_DIR / "output_type_classifier_metrics.json",
"w",
encoding="utf-8",
) as file:
json.dump(output_type_metrics, file, ensure_ascii=False, indent=2)
model_artifacts["output_type_classifier_metrics"] = str(
OUT_DIR / "output_type_classifier_metrics.json"
)
model_artifacts["output_type_classifier_report"] = str(
OUT_DIR / "output_type_classifier_report.csv"
)
model_artifacts["output_type_confusion_matrix"] = str(
OUT_DIR / "output_type_confusion_matrix.csv"
)
model_artifacts["output_type_top_tokens"] = str(
OUT_DIR / "output_type_top_tokens.csv"
)
else:
rprint(
"[yellow]Skipping output_type classifier because there are too few "
"classes or rows.[/yellow]"
)
output_type_metrics = {}
rprint(Panel.fit("[bold]Baseline 2: Predict tool_name from context using pure Python Naive Bayes[/bold]"))
tool_classifier_df = df[
df["output_type"].eq("tool_use")
& df["tool_name"].fillna("").astype(str).str.len().gt(0)
].copy()
if len(tool_classifier_df) >= 50 and tool_classifier_df["tool_name"].nunique() >= 2:
top_tools = tool_classifier_df["tool_name"].value_counts().head(12).index.tolist()
tool_classifier_df["tool_label"] = tool_classifier_df["tool_name"].where(
tool_classifier_df["tool_name"].isin(top_tools),
"__OTHER__",
)
y_tool = tool_classifier_df["tool_label"].astype(str).tolist()
X_tool_text = (
tool_classifier_df["context"]
.fillna("")
.astype(str)
.map(lambda text: text[:12000])
.tolist()
)
if len(set(y_tool)) >= 2:
train_indices, test_indices = stratified_train_test_indices(y_tool, test_size=0.2, seed=SEED)
X_train = [X_tool_text[i] for i in train_indices]
y_train = [y_tool[i] for i in train_indices]
X_test = [X_tool_text[i] for i in test_indices]
y_test = [y_tool[i] for i in test_indices]
tool_classifier = PureMultinomialNB(
max_features=20000,
min_df=2,
alpha=1.0,
)
tool_classifier.fit(X_train, y_train)
tool_predictions = tool_classifier.predict(X_test)
tool_metrics, tool_report_df = evaluate_predictions(y_test, tool_predictions)
tool_matrix_df = confusion_matrix_df(y_test, tool_predictions)
tool_metrics["train_rows"] = len(X_train)
tool_metrics["test_rows"] = len(X_test)
tool_metrics["vocab_size"] = len(tool_classifier.vocab)
rprint("[bold]Tool classifier report:[/bold]")
display(tool_report_df)
display(tool_matrix_df)
tool_report_df.to_csv(OUT_DIR / "tool_name_classifier_report.csv", index=False)
tool_matrix_df.to_csv(OUT_DIR / "tool_name_confusion_matrix.csv")
top_tool_token_records = []
for label in tool_classifier.labels:
for token, margin in tool_classifier.top_tokens_for_class(label, n=25):
top_tool_token_records.append(
{
"label": label,
"token": token,
"score_margin": margin,
}
)
pd.DataFrame(top_tool_token_records).to_csv(
OUT_DIR / "tool_name_top_tokens.csv",
index=False,
)
with open(
OUT_DIR / "tool_name_classifier_metrics.json",
"w",
encoding="utf-8",
) as file:
json.dump(tool_metrics, file, ensure_ascii=False, indent=2)
model_artifacts["tool_name_classifier_metrics"] = str(
OUT_DIR / "tool_name_classifier_metrics.json"
)
model_artifacts["tool_name_classifier_report"] = str(
OUT_DIR / "tool_name_classifier_report.csv"
)
model_artifacts["tool_name_confusion_matrix"] = str(
OUT_DIR / "tool_name_confusion_matrix.csv"
)
model_artifacts["tool_name_top_tokens"] = str(
OUT_DIR / "tool_name_top_tokens.csv"
)
else:
rprint("[yellow]Skipping tool classifier because labels collapsed to one class.[/yellow]")
tool_metrics = {}
else:
rprint(
"[yellow]Skipping tool classifier because there are too few tool-use "
"rows or tool classes.[/yellow]"
)
tool_metrics = {}
rprint(Panel.fit("[bold]Building simple keyword search helper[/bold]"))
def search_rows(keyword, limit=5, search_cols=("context", "cot", "completion", "text_payload")):
keyword = str(keyword).lower()
mask = pd.Series(False, index=df.index)
for column in search_cols:
mask = mask | (
df[column]
.fillna("")
.astype(str)
.str.lower()
.str.contains(re.escape(keyword), regex=True)
)
hits = df[mask].head(limit)
results = []
for _, row in hits.iterrows():
results.append(
{
"uid": row.get("uid"),
"session": row.get("session"),
"output_type": row.get("output_type"),
"tool_name": row.get("tool_name"),
"context_preview": preview_text(row.get("context"), 400),
"payload_preview": preview_text(row.get("text_payload"), 400),
}
)
return results
example_queries = [
"Bash",
"Write",
"browser",
"test",
"README",
]
search_demo = {
query: search_rows(query, limit=2)
for query in example_queries
}
with open(
OUT_DIR / "keyword_search_demo.json",
"w",
encoding="utf-8",
) as file:
json.dump(search_demo, file, ensure_ascii=False, indent=2)
rprint("[bold]Example keyword search results:[/bold]")
rprint(safe_json_dumps(search_demo, max_chars=5000))
summary = {
"dataset_id": DATASET_ID,
"flat_jsonl_filename": FLAT_JSONL_FILENAME,
"output_directory": str(OUT_DIR),
"repo_file_summary": file_summary,
"rows": int(len(df)),
"columns": list(df.columns),
"output_type_distribution": (
df["output_type"]
.fillna("missing")
.value_counts()
.to_dict()
),
"top_tools": (
df.loc[df["output_type"].eq("tool_use"), "tool_name"]
.replace("", "unknown")
.value_counts()
.head(20)
.to_dict()
),
"top_source_roots": (
df["source_root"]
.fillna("unknown")
.value_counts()
.head(20)
.to_dict()
),
"length_summary": {
column: {
"mean": float(df[column].mean()),
"median": float(df[column].median()),
"p90": float(df[column].quantile(0.90)),
"p95": float(df[column].quantile(0.95)),
"max": int(df[column].max()),
}
for column in [
"context_chars",
"cot_chars",
"completion_chars",
"text_payload_chars",
]
},
"possible_secret_rows": int(df["possible_secret_anywhere"].sum()),
"plots": plot_paths,
"model_artifacts": model_artifacts,
"safe_exports": {
"train": str(OUT_DIR / "fable5_no_cot_chat_train.jsonl"),
"validation": str(OUT_DIR / "fable5_no_cot_chat_validation.jsonl"),
"test": str(OUT_DIR / "fable5_no_cot_chat_test.jsonl"),
},
"analysis_files": {
"csv": str(OUT_DIR / "fable5_analysis_index.csv"),
"pickle": str(OUT_DIR / "fable5_analysis_index.pkl"),
"keyword_search_demo": str(OUT_DIR / "keyword_search_demo.json"),
},
}
with open(
OUT_DIR / "analysis_summary.json",
"w",
encoding="utf-8",
) as file:
json.dump(clean_for_json(summary), file, ensure_ascii=False, indent=2, default=str)
FENCE = chr(96) * 3
report_md = (
"# Fable 5 Traces Advanced Tutorial Report\n\n"
"## Dataset\n\n"
f"- Dataset: `{DATASET_ID}`\n"
f"- Flat JSONL: `{FLAT_JSONL_FILENAME}`\n"
f"- Rows loaded: `{len(df):,}`\n"
f"- Unique source sessions: `{df['session'].nunique(dropna=True):,}`\n"
f"- Unique models: `{df['model'].nunique(dropna=True):,}`\n\n"
"## Important safety note\n\n"
"This tutorial treats the dataset as agent telemetry. It previews and analyzes commands, "
"tool calls, file edits, and transcript text, but it never executes commands found inside "
"the traces.\n\n"
f"Potential secret-like patterns detected: `{int(df['possible_secret_anywhere'].sum()):,}` rows.\n"
"Exports redact common API-key/token-like patterns.\n\n"
"## Output type distribution\n\n"
f"{FENCE}json\n"
f"{json.dumps(clean_for_json(summary['output_type_distribution']), indent=2, ensure_ascii=False)}\n"
f"{FENCE}\n\n"
"## Top tools\n\n"
f"{FENCE}json\n"
f"{json.dumps(clean_for_json(summary['top_tools']), indent=2, ensure_ascii=False)}\n"
f"{FENCE}\n\n"
"## Saved files\n\n"
"- `analysis_summary.json`\n"
"- `fable5_analysis_index.csv`\n"
"- `fable5_analysis_index.pkl`\n"
"- `fable5_no_cot_chat_train.jsonl`\n"
"- `fable5_no_cot_chat_validation.jsonl`\n"
"- `fable5_no_cot_chat_test.jsonl`\n"
"- plot PNG files\n"
"- baseline classifier metrics, when enough rows/classes are available\n\n"
"## Recommended next steps\n\n"
"1. Inspect `fable5_no_cot_chat_train.jsonl` before any fine-tuning.\n"
"2. Keep the dataset license in mind before model training or redistribution.\n"
"3. Avoid training directly on raw terminal outputs without additional privacy and safety filtering.\n"
"4. Start with the no-CoT chat export unless your research explicitly requires reasoning-trace supervision.\n"
)
with open(
OUT_DIR / "REPORT.md",
"w",
encoding="utf-8",
) as file:
file.write(report_md)
rprint(
Panel.fit(
f"[bold green]Tutorial complete.[/bold green]\n\n"
f"Artifacts saved in:\n{OUT_DIR}\n\n"
f"Key files:\n"
f"- {OUT_DIR / 'REPORT.md'}\n"
f"- {OUT_DIR / 'analysis_summary.json'}\n"
f"- {OUT_DIR / 'fable5_no_cot_chat_train.jsonl'}\n"
f"- {OUT_DIR / 'fable5_analysis_index.csv'}",
title="Done",
)
)
display(
pd.DataFrame(
{
"artifact": [
"Report",
"Summary JSON",
"No-CoT train export",
"No-CoT validation export",
"No-CoT test export",
"Analysis CSV",
"Analysis pickle",
"Keyword search demo",
],
"path": [
str(OUT_DIR / "REPORT.md"),
str(OUT_DIR / "analysis_summary.json"),
str(OUT_DIR / "fable5_no_cot_chat_train.jsonl"),
str(OUT_DIR / "fable5_no_cot_chat_validation.jsonl"),
str(OUT_DIR / "fable5_no_cot_chat_test.jsonl"),
str(OUT_DIR / "fable5_analysis_index.csv"),
str(OUT_DIR / "fable5_analysis_index.pkl"),
str(OUT_DIR / "keyword_search_demo.json"),
],
}
)
)
