inspect-ai

分类: 调研与分析 | 上传者: UKGovernmentBEIS | 下载: 0 | 版本: v1.0（最新）
Analyze Inspect AI evaluation logs, understand EvalLog structure, extract samples, events, and scoring data using dataframes
更新日志: Source: GitHub https://github.com/UKGovernmentBEIS/sandbox_escape_bench
目录结构

当前层级: transcript-analysis/skills/inspect-ai/
📄 SKILL.md 12.8 KB
SKILL.md

---
name: inspect-ai
description: Analyze Inspect AI evaluation logs, understand EvalLog structure, extract samples, events, and scoring data using dataframes
user-invocable: false
---

# Inspect AI Log Analysis Reference

Use this knowledge when working with Inspect AI evaluation logs (`.eval` or `.json` files).

## File Format

`.eval` files are binary (compressed) format containing JSON data. They are essentially zip archives. To read them programmatically, use the Inspect Python API.

## Core Data Structures

### EvalLog (Top-level)

```python
class EvalLog:
    version: int                    # File format version (currently 2)
    status: str                     # "started", "success", or "error"
    eval: EvalSpec                  # Task, model, creation time
    plan: EvalPlan                  # Solvers and generation config
    results: EvalResults            # Aggregate scorer metrics
    stats: EvalStats                # Token usage statistics
    error: EvalError | None         # Error details if status="error"
    samples: list[EvalSample]       # Individual sample records
    reductions: list[EvalSampleReduction]  # Multi-epoch reductions
```

**Always check `log.status == "success"` before analyzing results.**

### EvalSample (Per-sample data)

```python
class EvalSample:
    id: int | str                   # Unique sample identifier
    epoch: int                      # Epoch number (for multi-epoch runs)
    input: str | list[ChatMessage]  # The prompt/task given to model
    target: str | list[str]         # Expected answer(s)
    choices: list[str] | None       # Multiple choice options if applicable

    # Execution results
    messages: list[ChatMessage]     # Full conversation history
    output: ModelOutput             # Final model output
    scores: dict[str, Score] | None # Scoring results by scorer name

    # Events and state
    events: list[Event]             # Complete transcript of all events
    store: dict[str, Any]           # State at end of execution
    attachments: dict[str, str]     # Referenced content (images, etc.)

    # Metadata
    metadata: dict[str, Any]        # Custom key-value pairs from dataset
    sandbox: SandboxEnvironmentSpec | None  # Sandbox config
    files: list[str] | None         # Files provided to sandbox
    setup: str | None               # Setup script run in sandbox

    # Timing
    started_at: datetime | None
    completed_at: datetime | None
    total_time: float | None        # Wall clock time
    working_time: float | None      # Active processing time

    # Token usage
    model_usage: dict[str, ModelUsage]  # Tokens by model

    # Error/limit info
    error: EvalError | None         # Error that halted sample
    error_retries: list[EvalError] | None  # Retried errors
    limit: EvalSampleLimit | None   # Limit that halted sample (context/time/message/token)
```

### Event Types (for behavioral analysis)

Events are the core of behavioral analysis. Each sample has an `events` list containing:

```python
Event = Union[
    SampleInitEvent,    # Sample initialization
    SampleLimitEvent,   # Limit reached
    SandboxEvent,       # Sandbox operations (exec, read_file, write_file)
    StateEvent,         # State changes
    StoreEvent,         # Store updates
    ModelEvent,         # LLM API calls
    ToolEvent,          # Tool invocations
    ApprovalEvent,      # Human approval events
    InputEvent,         # User input
    ScoreEvent,         # Scoring events
    ScoreEditEvent,     # Score modifications
    ErrorEvent,         # Errors
    LoggerEvent,        # Log messages
    InfoEvent,          # Info messages
    SpanBeginEvent,     # Span start (for timing)
    SpanEndEvent,       # Span end
    StepEvent,          # Solver step events
    SubtaskEvent,       # Subtask events
]
```

#### ModelEvent (LLM calls)

```python
class ModelEvent:
    event: "model"
    model: str                      # Model name
    input: list[ChatMessage]        # Messages sent to model
    tools: list[ToolInfo]           # Available tools
    tool_choice: ToolChoice         # Tool selection directive
    config: GenerateConfig          # Generation parameters
    output: ModelOutput             # Model response
    retries: int | None             # API retries
    error: str | None               # Error if failed
    cache: "read" | "write" | None  # Cache hit/miss
    timestamp: datetime             # When call started
    completed: datetime | None      # When call finished
    working_time: float | None      # Processing time
```

#### ToolEvent (Tool calls)

```python
class ToolEvent:
    event: "tool"
    id: str                         # Unique tool call ID
    function: str                   # Tool/function name
    arguments: dict[str, JsonValue] # Arguments passed
    result: ToolResult              # Return value
    error: ToolCallError | None     # Error if failed
    truncated: tuple[int, int] | None  # If output was truncated
    timestamp: datetime             # When call started
    completed: datetime | None      # When call finished
    working_time: float | None      # Processing time
    agent: str | None               # Agent name if handoff
    failed: bool | None             # Hard failure flag
```

### Score

```python
class Score:
    value: float | str | int | bool | list  # The score value
    answer: str | None              # Model's answer extracted
    explanation: str | None         # Explanation of score
    metadata: dict[str, Any] | None # Additional scoring metadata
    history: list[ScoreEdit]        # Edit history (history[0] = original)
```

---

## Dataframe API (Primary Analysis Method)

The `inspect_ai.analysis` module provides functions to convert logs into Pandas dataframes.

### evals_df() - One row per evaluation

```python
from inspect_ai.analysis import evals_df

df = evals_df("logs")  # Read all logs in directory
df = evals_df(["path/to/file1.eval", "path/to/file2.eval"])  # Specific files
```

**Default columns (~51):**
- `eval_id` - Unique evaluation identifier
- `log` - URI of source file
- `task`, `task_version`, `task_file`, `task_arg_*` - Task info
- `model`, `model_args`, `generate_config_*` - Model info
- `status`, `error` - Completion status
- `score_<scorer>_<metric>` - All scores expanded as columns
- `samples_completed`, `samples_total`
- `created`, `git_commit`, `tags`, `metadata_*`

**Pre-built column groups:**
```python
from inspect_ai.analysis import (
    EvalInfo,      # created, tags, metadata, git
    EvalTask,      # task name, file, args, solver
    EvalModel,     # model name, args, generation config
    EvalDataset,   # dataset name, location, sample IDs
    EvalConfig,    # epochs, approval, sample limits
    EvalResults,   # status, errors, samples completed
    EvalScores,    # all scores as separate columns
    EvalColumns,   # all of the above (~50 columns)
)
```

### samples_df() - One row per sample

```python
from inspect_ai.analysis import samples_df, SampleSummary, SampleScores, SampleMessages

# Fast read (summaries only, 12 columns)
df = samples_df("logs")

# With detailed scores
df = samples_df("logs", columns=SampleSummary + SampleScores)

# With message content (slower, loads full samples)
df = samples_df("logs", columns=SampleSummary + SampleMessages)
```

**SampleSummary columns (default, 12 columns):**
- `sample_id` - Globally unique identifier
- `eval_id` - Links to evaluation
- `id`, `epoch` - Sample ID within eval and epoch number
- `input`, `target` - Task input and expected output
- `metadata_*` - Expanded metadata dictionary
- `score_*` - Score values only
- `model_usage` - Token counts
- `total_time`, `working_time` - Timing data
- `error`, `limit`, `retries` - Failure info
- `log` - Source file URI

**SampleScores adds:**
- Score answer, explanation, metadata

**SampleMessages adds:**
- Full message content (requires loading full sample)

### messages_df() - One row per message

```python
from inspect_ai.analysis import messages_df

# All messages
df = messages_df("logs")

# Filter by role
df = messages_df("logs", filter=["assistant"])
df = messages_df("logs", filter=["user", "assistant"])

# Custom filter function
df = messages_df("logs", filter=lambda msg: "error" in msg.content.lower())
```

**Default columns:**
- `sample_id`, `eval_id` - Links to sample and evaluation
- `event_id` - Unique message identifier
- `role` - user, assistant, system, tool
- `content` - Message text
- `source` - Origin of message
- `tool_calls` - Formatted function calls
- `tool_call_id`, `tool_call_function`, `tool_call_error`
- `log` - Source file URI

### events_df() - One row per event

```python
from inspect_ai.analysis import (
    events_df,
    EventInfo,           # event type, span ID
    EventTiming,         # start/end times
    ModelEventColumns,   # model event data
    ToolEventColumns,    # tool event data
)

# Must specify columns (events are heterogeneous)
df = events_df("logs", columns=EventInfo + EventTiming)

# Filter to specific event types
df = events_df("logs", columns=EventInfo + ToolEventColumns,
               filter=lambda e: e.event == "tool")
```

**EventInfo columns:**
- `event_type` - Type of event (model, tool, sandbox, etc.)
- `span_id` - Span identifier for grouping

**EventTiming columns:**
- `timestamp` - When event started
- `completed` - When event finished
- `working_time` - Active processing time

---

## Joining Dataframes

Use `eval_id` and `sample_id` to join across dataframes:

```python
# Join evals with samples
merged = samples.merge(evals, on='eval_id')

# Join samples with messages
merged = messages.merge(samples, on='sample_id')

# DuckDB integration
import duckdb
con = duckdb.connect()
con.register('evals', evals_df("logs"))
con.register('samples', samples_df("logs"))
con.execute("""
    SELECT e.model, AVG(s.score_accuracy)
    FROM samples s JOIN evals e ON s.eval_id = e.eval_id
    GROUP BY e.model
""")
```

---

## Data Preparation Functions

```python
from inspect_ai.analysis import prepare, model_info, task_info, frontier

# Add model metadata columns
df = prepare(df, model_info())
# Adds: model_organization_name, model_display_name, model_snapshot,
#       model_release_date, model_knowledge_cutoff_date

# Map task names to display names
df = prepare(df, task_info({"gpqa_diamond": "GPQA Diamond"}))

# Add frontier indicator (requires model_info first)
df = prepare(df, frontier())
# Adds boolean column: was model top-scoring at release date?
```

---

## Low-level Log Reading API

```python
from inspect_ai.log import (
    read_eval_log,
    read_eval_log_sample,
    read_eval_log_samples,
    read_eval_log_sample_summaries,
    list_eval_logs,
)

# Read complete log
log = read_eval_log("path/to/file.eval")

# Read just header (no samples) - fast for large files
log = read_eval_log("path/to/file.eval", header_only=True)

# Stream samples one at a time (memory efficient)
for sample in read_eval_log_samples("path/to/file.eval"):
    process(sample)

# Get lightweight summaries for filtering
summaries = read_eval_log_sample_summaries("path/to/file.eval")

# Read specific sample
sample = read_eval_log_sample("path/to/file.eval", id="sample_id", epoch=1)

# List all logs in directory
logs = list_eval_logs("./logs", recursive=True)
```

---

## CLI Commands

```bash
# List logs with filtering
inspect log list --json
inspect log list --status success

# Dump log as JSON
inspect log dump path/to/file.eval

# Convert between formats
inspect log convert file.json --to eval --output-dir ./converted
```

---

## Common Analysis Patterns

### QA Verification

```python
# Check all evaluations completed successfully
evals = evals_df("logs")
failed = evals[evals['status'] != 'success']

# Find samples with errors
samples = samples_df("logs")
errored = samples[samples['error'].notna()]

# Find samples that hit limits
limited = samples[samples['limit'].notna()]
```

### Behavioral Analysis

```python
# Get tool usage patterns
events = events_df("logs", columns=EventInfo + ToolEventColumns,
                   filter=lambda e: e.event == "tool")
tool_counts = events.groupby(['eval_id', 'function']).size()

# Analyze message patterns
messages = messages_df("logs")
msg_counts = messages.groupby(['eval_id', 'role']).size().unstack()

# Compare successful vs failed attempts
samples = samples_df("logs")
successful = samples[samples['score_accuracy'] == 1.0]
failed = samples[samples['score_accuracy'] == 0.0]
```

### Cross-model Comparison

```python
evals = evals_df("logs")
by_model = evals.groupby('model').agg({
    'score_accuracy_mean': 'mean',
    'samples_completed': 'sum'
})
```

---

## Performance Tips

- Use `SampleSummary` (default) for fast reads - only loads headers
- Use `parallel=True` for large datasets: `samples_df("logs", parallel=True)`
- Use `header_only=True` with `read_eval_log()` when you don't need samples
- Stream with `read_eval_log_samples()` for memory-constrained environments
- Use `strict=False` to get partial results: `df, errors = evals_df("logs", strict=False)`
登录后下载/点赞/收藏 ❤ 2 | ★ 0
inspect-ai

目录结构

SKILL.md

举报内容

提示
