trl/references/reliability_principles.md

Reliability Principles for Training Jobs

These principles are derived from real production failures and successful fixes. Following them prevents common failure modes and ensures reliable job execution.

Principle 1: Always Verify Before Use

Rule: Never assume repos, datasets, or resources exist. Verify with tools first.

What It Prevents

• Non-existent datasets - Jobs fail immediately when dataset doesn't exist
• Typos in names - Simple mistakes like "argilla-dpo-mix-7k" vs "ultrafeedback_binarized"
• Incorrect paths - Old or moved repos, renamed files
• Missing dependencies - Undocumented requirements

How to Apply

Before submitting ANY job:

# Verify dataset exists
dataset_search({"query": "dataset-name", "author": "author-name", "limit": 5})
hub_repo_details(["author/dataset-name"], repo_type="dataset")

# Verify model exists
hub_repo_details(["org/model-name"], repo_type="model")

# Check script/file paths (for URL-based scripts)
# Verify before using: https://github.com/user/repo/blob/main/script.py

Examples that would have caught errors:

# ❌ WRONG: Assumed dataset exists
hf_jobs("uv", {
    "script": """...""",
    "env": {"DATASET": "trl-lib/argilla-dpo-mix-7k"}  # Doesn't exist!
})

# ✅ CORRECT: Verify first
dataset_search({"query": "argilla dpo", "author": "trl-lib"})
# Would show: "trl-lib/ultrafeedback_binarized" is the correct name

hub_repo_details(["trl-lib/ultrafeedback_binarized"], repo_type="dataset")
# Confirms it exists before using

Implementation Checklist

• Check dataset exists before training
• Verify base model exists before fine-tuning
• Confirm adapter model exists before GGUF conversion
• Test script URLs are valid before submitting
• Validate file paths in repositories
• Check for recent updates/renames of resources

Time cost: 5-10 seconds
Time saved: Hours of failed job time + debugging

---

Principle 2: Prioritize Reliability Over Performance

Rule: Default to what is most likely to succeed, not what is theoretically fastest.

What It Prevents

• Hardware incompatibilities - Features that fail on certain GPUs
• Unstable optimizations - Speed-ups that cause crashes
• Complex configurations - More failure points
• Build system issues - Unreliable compilation methods

How to Apply

Choose reliability:

# ❌ RISKY: Aggressive optimization that may fail
SFTConfig(
    torch_compile=True,  # Can fail on T4, A10G GPUs
    optim="adamw_bnb_8bit",  # Requires specific setup
    fp16=False,  # May cause training instability
    ...
)

# ✅ SAFE: Proven defaults
SFTConfig(
    # torch_compile=True,  # Commented with note: "Enable on H100 for 20% speedup"
    optim="adamw_torch",  # Standard, always works
    fp16=True,  # Stable and fast
    ...
)

For build processes:

# ❌ UNRELIABLE: Uses make (platform-dependent)
subprocess.run(["make", "-C", "/tmp/llama.cpp", "llama-quantize"], check=True)

# ✅ RELIABLE: Uses CMake (consistent, documented)
subprocess.run([
    "cmake", "-B", "/tmp/llama.cpp/build", "-S", "/tmp/llama.cpp",
    "-DGGML_CUDA=OFF"  # Disable CUDA for faster, more reliable build
], check=True)

subprocess.run([
    "cmake", "--build", "/tmp/llama.cpp/build",
    "--target", "llama-quantize", "-j", "4"
], check=True)

Real-World Example

The torch.compile failure:

• Added for "20% speedup" on H100
• Failed fatally on T4-medium with cryptic error
• Misdiagnosed as dataset issue (cost hours)
• Fix: Disable by default, add as optional comment

Result: Reliability > 20% performance gain

Implementation Checklist

• Use proven, standard configurations by default
• Comment out performance optimizations with hardware notes
• Use stable build systems (CMake > make)
• Test on target hardware before production
• Document known incompatibilities
• Provide "safe" and "fast" variants when needed

Performance loss: 10-20% in best case
Reliability gain: 95%+ success rate vs 60-70%

---

Principle 3: Create Atomic, Self-Contained Scripts

Rule: Scripts should work as complete, independent units. Don't remove parts to "simplify."

What It Prevents

• Missing dependencies - Removed "unnecessary" packages that are actually required
• Incomplete processes - Skipped steps that seem redundant
• Environment assumptions - Scripts that need pre-setup
• Partial failures - Some parts work, others fail silently

How to Apply

Complete dependency specifications:

# ❌ INCOMPLETE: "Simplified" by removing dependencies
# /// script
# dependencies = [
#     "transformers",
#     "peft",
#     "torch",
# ]
# ///

# ✅ COMPLETE: All dependencies explicit
# /// script
# dependencies = [
#     "transformers>=4.36.0",
#     "peft>=0.7.0",
#     "torch>=2.0.0",
#     "accelerate>=0.24.0",
#     "huggingface_hub>=0.20.0",
#     "sentencepiece>=0.1.99",  # Required for tokenizers
#     "protobuf>=3.20.0",        # Required for tokenizers
#     "numpy",
#     "gguf",
# ]
# ///

Complete build processes:

# ❌ INCOMPLETE: Assumes build tools exist
subprocess.run(["git", "clone", "https://github.com/ggerganov/llama.cpp.git", "/tmp/llama.cpp"])
subprocess.run(["make", "-C", "/tmp/llama.cpp", "llama-quantize"])  # FAILS: no gcc/make

# ✅ COMPLETE: Installs all requirements
subprocess.run(["apt-get", "update", "-qq"], check=True)
subprocess.run(["apt-get", "install", "-y", "-qq", "build-essential", "cmake"], check=True)
subprocess.run(["git", "clone", "https://github.com/ggerganov/llama.cpp.git", "/tmp/llama.cpp"])
# ... then build

Real-World Example

The sentencepiece failure:

• Original script had it: worked fine
• "Simplified" version removed it: "doesn't look necessary"
• GGUF conversion failed silently - tokenizer couldn't convert
• Hard to debug: no obvious error message
• Fix: Restore all original dependencies

Result: Don't remove dependencies without thorough testing

Implementation Checklist

• All dependencies in PEP 723 header with version pins
• All system packages installed by script
• No assumptions about pre-existing environment
• No "optional" steps that are actually required
• Test scripts in clean environment
• Document why each dependency is needed

Complexity: Slightly longer scripts
Reliability: Scripts "just work" every time

---

Principle 4: Provide Clear Error Context

Rule: When things fail, make it obvious what went wrong and how to fix it.

How to Apply

Wrap subprocess calls:

# ❌ UNCLEAR: Silent failure
subprocess.run([...], check=True, capture_output=True)

# ✅ CLEAR: Shows what failed
try:
    result = subprocess.run(
        [...],
        check=True,
        capture_output=True,
        text=True
    )
    print(result.stdout)
    if result.stderr:
        print("Warnings:", result.stderr)
except subprocess.CalledProcessError as e:
    print(f"❌ Command failed!")
    print("STDOUT:", e.stdout)
    print("STDERR:", e.stderr)
    raise

Validate inputs:

# ❌ UNCLEAR: Fails later with cryptic error
model = load_model(MODEL_NAME)

# ✅ CLEAR: Fails fast with clear message
if not MODEL_NAME:
    raise ValueError("MODEL_NAME environment variable not set!")

print(f"Loading model: {MODEL_NAME}")
try:
    model = load_model(MODEL_NAME)
    print(f"✅ Model loaded successfully")
except Exception as e:
    print(f"❌ Failed to load model: {MODEL_NAME}")
    print(f"Error: {e}")
    print("Hint: Check that model exists on Hub")
    raise

Implementation Checklist

• Wrap external calls with try/except
• Print stdout/stderr on failure
• Validate environment variables early
• Add progress indicators (✅, ❌, 🔄)
• Include hints for common failures
• Log configuration at start

---

Principle 5: Test the Happy Path on Known-Good Inputs

Rule: Before using new code in production, test with inputs you know work.

How to Apply

Known-good test inputs:

# For training
TEST_DATASET = "trl-lib/Capybara"  # Small, well-formatted, widely used
TEST_MODEL = "Qwen/Qwen2.5-0.5B"  # Small, fast, reliable

# For GGUF conversion
TEST_ADAPTER = "evalstate/qwen-capybara-medium"  # Known working model
TEST_BASE = "Qwen/Qwen2.5-0.5B"  # Compatible base

Testing workflow:

1. Test with known-good inputs first
2. If that works, try production inputs
3. If production fails, you know it's the inputs (not code)
4. Isolate the difference

Implementation Checklist

• Maintain list of known-good test models/datasets
• Test new scripts with test inputs first
• Document what makes inputs "good"
• Keep test jobs cheap (small models, short timeouts)
• Only move to production after test succeeds

Time cost: 5-10 minutes for test run
Debugging time saved: Hours

---

Summary: The Reliability Checklist

Before submitting ANY job:

Pre-Flight Checks

• Verified all repos/datasets exist (hub_repo_details)
• Tested with known-good inputs if new code
• Using proven hardware/configuration
• Included all dependencies in PEP 723 header
• Installed system requirements (build tools, etc.)
• Set appropriate timeout (not default 30m)
• Configured Hub push with HF_TOKEN
• Added clear error handling

Script Quality

• Self-contained (no external setup needed)
• Complete dependencies listed
• Build tools installed by script
• Progress indicators included
• Error messages are clear
• Configuration logged at start

Job Configuration

• Timeout > expected runtime + 30% buffer
• Hardware appropriate for model size
• Secrets include HF_TOKEN
• Environment variables set correctly
• Cost estimated and acceptable

Following these principles transforms job success rate from ~60-70% to ~95%+

---

When Principles Conflict

Sometimes reliability and performance conflict. Here's how to choose:

Scenario	Choose	Rationale
Demo/test	Reliability	Fast failure is worse than slow success
Production (first run)	Reliability	Prove it works before optimizing
Production (proven)	Performance	Safe to optimize after validation
Time-critical	Reliability	Failures cause more delay than slow runs
Cost-critical	Balanced	Test with small model, then optimize

General rule: Reliability first, optimize second.

---

Further Reading

• troubleshooting.md - Common issues and fixes
• training_patterns.md - Proven training configurations
• gguf_conversion.md - Production GGUF workflow