Saurabhs/view tool rl

.claude/settings.local.json

@@ -0,0 +1,30 @@
+{
+  "permissions": {
+    "allow": [
+      "Bash(chmod:*)",
+      "Bash(make:*)",
+      "Bash(python test:*)",
+      "Bash(python:*)",
+      "Bash(ls:*)",
+      "Bash(rm:*)",
+      "Bash(uv run ruff check:*)",
+      "Bash(uv run pytest:*)",
+      "Bash(uv run:*)",
+      "Bash(ldd:*)",
+      "Bash(uv pip show:*)",
+      "Bash(uv sync:*)",
+      "Bash(uv pip uninstall:*)",
+      "Bash(curl:*)",
+      "Bash(grep:*)",
+      "Bash(bash:*)",
+      "Bash(sed:*)",
+      "Bash(awk:*)",
+      "WebSearch",
+      "WebFetch(domain:docs.ray.io)",
+      "Bash(ss:*)",
+      "Bash(uv pip:*)"
+    ],
+    "deny": [],
+    "defaultMode": "acceptEdits"
+  }
+}

.gitignore

@@ -159,3 +159,5 @@ dmypy.json
 cache/
 local_dataset_cache/
 scratch/
+open_instruct/code_utils/testbed/
+open_instruct/code_utils/repos/

build_and_push_image.sh

@@ -0,0 +1,13 @@
+#!/bin/bash
+image_name=open-coding-agent
+
+# Build and push the Docker image to Beaker
+docker build . --build-arg UV_CACHE_DIR=$UV_CACHE_DIR -t $image_name
+
+beaker_user=$(beaker account whoami --format json | jq -r '.[0].name')
+
+# Use '|| true' to prevent script from exiting if image doesn't exist to delete.
+beaker image delete $beaker_user/$image_name || true
+
+# Create the image in the same workspace used for jobs.
+beaker image create $image_name -n $image_name -w ai2/$beaker_user

coding-agent/CODE_SEARCH_DATASETS_README.md

@@ -0,0 +1,255 @@
+# Code Search Datasets Documentation
+
+## Overview
+
+This document describes the two HuggingFace datasets created for code search tasks:
+
+1. **Multi-Step Tool Dataset** (`HF_OUTPUT_MULTI_STEP_TOOL`) - For models that use the CodeSearchTool with multiple interactions
+2. **Single-Turn Dataset** - For models that output a single view call to find buggy code
+
+## Dataset Creation
+
+### Script: `create_code_search_datasets.py`
+
+The main script that transforms raw coding-agent data into structured datasets.
+
+**Usage:**
+```bash
+# Process all data files
+python create_code_search_datasets.py \
+    --data-dir coding-agent/data \
+    --output-dir code_search_datasets
+
+# Test with sample.json only
+python create_code_search_datasets.py \
+    --single-file-test \
+    --output-dir test_datasets
+
+# Push to HuggingFace Hub
+python create_code_search_datasets.py \
+    --push-to-hub \
+    --hub-org your-org-name
+```
+
+## Dataset Formats
+
+### 1. Multi-Step Tool Dataset
+
+**Purpose:** Training models to use CodeSearchTool for exploring repositories with multiple tool calls.
+
+**Schema:**
+```python
+{
+    "instance_id": str,           # Unique identifier (e.g., "starlette_10457")
+    "messages": List[Dict],       # Full conversation with system, user, assistant messages
+    "buggy_info": Dict,          # Information about the bug location
+    "num_turns": int,            # Total number of conversation turns
+    "tool_calls_made": int       # Number of tool calls in the conversation
+}
+```
+
+**Message Format:**
+- Uses the full conversation history including tool responses
+- Includes system prompts with tool descriptions
+- Multiple assistant responses with tool calls
+
+**Example Entry:**
+```python
+{
+    "instance_id": "starlette_10457",
+    "messages": [
+        {"role": "system", "content": "You are a helpful assistant..."},
+        {"role": "user", "content": "Find the bug in Config.__call__..."},
+        {"role": "assistant", "content": "<tool_call>{...}</tool_call>"},
+        {"role": "user", "content": "<tool_response>...</tool_response>"},
+        ...
+    ],
+    "buggy_info": {
+        "file_path": "/testbed/starlette/config.py",
+        "buggy_line": 130,
+        "view_range": [121, 138]
+    },
+    "num_turns": 32,
+    "tool_calls_made": 15
+}
+```
+
+### 2. Single-Turn Dataset
+
+**Purpose:** Training models to directly identify and view the buggy file in a single response.
+
+**Schema:**
+```python
+{
+    "instance_id": str,          # Unique identifier
+    "messages": List[Dict],      # Single-turn conversation (system, user, assistant)
+    "buggy_file": str,          # Path to the file containing the bug
+    "buggy_line": int,          # Line number of the bug (-1 if unknown)
+    "bug_description": str       # Truncated description of the bug
+}
+```
+
+**Message Format:**
+- Simplified single-turn format
+- System prompt explains the task
+- User provides bug description
+- Assistant responds with a single view tool call
+
+**Example Entry:**
+```python
+{
+    "instance_id": "starlette_10457",
+    "messages": [
+        {
+            "role": "system", 
+            "content": "You are a code search assistant..."
+        },
+        {
+            "role": "user",
+            "content": "Repository location: /testbed\n\nBug description:\n..."
+        },
+        {
+            "role": "assistant",
+            "content": "I'll examine the file...\n\n<tool_call>{...}</tool_call>"
+        }
+    ],
+    "buggy_file": "/testbed/starlette/config.py",
+    "buggy_line": 130,
+    "bug_description": "Possible bug in Config.__call__..."
+}
+```
+
+## Integration with Existing Components
+
+### 1. CodeViewTool (`tool_vllm.py`)
+
+The `CodeViewTool` class processes tool calls from model outputs:
+
+```python
+from open_instruct.tool_utils.tool_vllm import CodeViewTool
+
+tool = CodeViewTool(
+    api_endpoint="http://localhost:1234",
+    repo_name="repository-name",
+    start_str="<tool_call>",
+    end_str="</tool_call>"
+)
+
+# Process model output containing tool calls
+result = tool(model_output)
+```
+
+### 2. CodeSearchVerifier (`ground_truth_utils.py`)
+
+The `CodeSearchVerifier` evaluates model predictions:
+
+```python
+from open_instruct.ground_truth_utils import CodeSearchVerifier, CodeVerifierConfig
+
+config = CodeVerifierConfig(
+    code_api_url="http://localhost:1234",
+    code_max_execution_time=5.0,
+    code_pass_rate_reward_threshold=0.5,
+    code_apply_perf_penalty=True
+)
+
+verifier = CodeSearchVerifier(config)
+
+# Evaluate prediction
+result = verifier(
+    tokenized_prediction=[],
+    prediction=model_output,
+    label=expected_files,
+    query=original_query
+)
+```
+
+### 3. View File API (`api.py`)
+
+The API endpoint for viewing repository files:
+
+```python
+# POST /view_file
+{
+    "repo_name": "cool-RR/PySnooper",
+    "path": "pysnooper/pycompat.py",
+    "view_range": [86, 88],  # Optional
+    "base_commit": "abc123"  # Optional
+}
+
+# Response
+{
+    "content": "file content here...",
+    "repo_path": "/path/to/cloned/repo"
+}
+```
+
+## Loading and Using the Datasets
+
+```python
+from datasets import load_from_disk
+
+# Load datasets
+multi_step = load_from_disk('code_search_datasets/multi_step_tool_dataset')
+single_turn = load_from_disk('code_search_datasets/single_turn_dataset')
+
+# Access samples
+for sample in multi_step:
+    print(f"Instance: {sample['instance_id']}")
+    print(f"Tool calls: {sample['tool_calls_made']}")
+
+for sample in single_turn:
+    print(f"Bug file: {sample['buggy_file']}")
+    print(f"Bug line: {sample['buggy_line']}")
+```
+
+## Training Considerations
+
+### For Multi-Step Models:
+- Use the full conversation history
+- Train to generate appropriate tool calls based on context
+- Handle tool responses and continue reasoning
+
+### For Single-Turn Models:
+- Focus on extracting key information from bug descriptions
+- Generate precise file paths and view ranges
+- Optimize for accuracy in first attempt
+
+## Evaluation Metrics
+
+The CodeSearchVerifier provides several metrics:
+- **File Match Score**: Whether the correct file was viewed
+- **Line Coverage**: Whether the buggy line was within the view range
+- **Efficiency Penalty**: Penalty for viewing unnecessary files
+- **Response Time**: Time taken to identify the bug
+
+## ChatML Template Support
+
+Both datasets support the ChatML template format implemented in `dataset_transformation.py`:
+
+```python
+from open_instruct.dataset_transformation import TokenizerConfig
+
+tc = TokenizerConfig(
+    tokenizer_name_or_path="your-model",
+    chat_template_name="chatml"  # Use ChatML format
+)
+```
+
+This formats messages as:
+```
+<|im_start|>system
+You are a code search assistant.<|im_end|>
+<|im_start|>user
+Find the bug in this code.<|im_end|>
+<|im_start|>assistant
+I'll search for the bug.<|im_end|>
+```
+
+## Future Enhancements
+
+1. **Multi-repository support**: Extend to handle bugs across multiple repositories
+2. **Enhanced scoring**: More sophisticated evaluation metrics for partial matches
+3. **Difficulty levels**: Categorize bugs by complexity
+4. **Cross-file bugs**: Support bugs that span multiple files
+5. **Test generation**: Include test cases that reproduce the bugs