An Empirical Study of Path Feasibility Queries
Paper • 1302.4798 • Published
AI & ML interests
code LLMs, static analysis, software composition analysis, vulnerability remediation, application security
Recent Activity
View all activity
Papers
Patched MOA: optimizing inference for diverse software development tasks
Patched RTC: evaluating LLMs for diverse software development tasks
codelion
authored 4
papers 3 days ago
Post
419
SPROG-9M — a 9.37M parameter model trained from scratch to solve GSM8K-style math without using an LLM at inference.
The model, codelion/sprog-9m, predicts symbolic programs over number slots, then a deterministic executor does the arithmetic. With a simple verifier, it reaches ~11.8% on GSM8K test.
We also released the dataset: codelion/gsm8k-synth, 117K validated synthetic GSM8K-style problems.
Tiny model, no pretraining, no LLM at inference, runs on a laptop.
The model, codelion/sprog-9m, predicts symbolic programs over number slots, then a deterministic executor does the arithmetic. With a simple verifier, it reaches ~11.8% on GSM8K test.
We also released the dataset: codelion/gsm8k-synth, 117K validated synthetic GSM8K-style problems.
Tiny model, no pretraining, no LLM at inference, runs on a laptop.
Post
3249
Inspired by the Nemotron Diffusion recipe, check out dhara-250m: a 250M experimental language model that supports three decoding modes from one set of weights: autoregressive, block-diffusion, and self-speculation.
It is small, easy to try, and meant for exploring diffusion-style decoding and latency tradeoffs in compact LMs.
Model: codelion/dhara-250m
Try the chat demo here: codelion/dhara-chat
It is small, easy to try, and meant for exploring diffusion-style decoding and latency tradeoffs in compact LMs.
Model: codelion/dhara-250m
Try the chat demo here: codelion/dhara-chat
- 3 replies
Post
3434
Scaling Pedagogical Pre-training to 10 Billion Tokens
New blog post exploring what happens when you take optimal data mixing insights and scale up the data generation itself.
We built Sutra, a multi-stage framework for generating pedagogical pre-training data guided by a knowledge graph of ~2,000 concepts across 9 domains. The pipeline includes structured content generation, six-dimension quality evaluation, diversity management across 20 content styles, and a cleaning stage to prevent collapse.
The result is codelion/sutra-10B, a 10.2 billion token pedagogical dataset with rich metadata (domain, complexity, prerequisites, quality scores) on every entry.
We trained codelion/SmolLM2-70M on it for 3 full epochs (30.6B tokens) on a single A10 GPU in ~78 hours.
Key finding: perplexity kept improving across epochs, but benchmark gains plateaued fast. At 70M parameters, the model hits a representational ceiling that more data alone can't break through.
Full writeup with comparisons against 7 other datasets, detailed benchmark breakdowns, and connections to recent work on synthetic data scaling, curriculum learning, and data mixing laws: https://huggingface.co/blog/codelion/scaling-pedagogical-pretraining-10-billion-tokens
All datasets at multiple scales (10M, 100M, 1B, 10B) plus seed concepts and an SFT variant are in the Sutra Pedagogical Datasets collection.
New blog post exploring what happens when you take optimal data mixing insights and scale up the data generation itself.
We built Sutra, a multi-stage framework for generating pedagogical pre-training data guided by a knowledge graph of ~2,000 concepts across 9 domains. The pipeline includes structured content generation, six-dimension quality evaluation, diversity management across 20 content styles, and a cleaning stage to prevent collapse.
The result is codelion/sutra-10B, a 10.2 billion token pedagogical dataset with rich metadata (domain, complexity, prerequisites, quality scores) on every entry.
We trained codelion/SmolLM2-70M on it for 3 full epochs (30.6B tokens) on a single A10 GPU in ~78 hours.
Key finding: perplexity kept improving across epochs, but benchmark gains plateaued fast. At 70M parameters, the model hits a representational ceiling that more data alone can't break through.
Full writeup with comparisons against 7 other datasets, detailed benchmark breakdowns, and connections to recent work on synthetic data scaling, curriculum learning, and data mixing laws: https://huggingface.co/blog/codelion/scaling-pedagogical-pretraining-10-billion-tokens
All datasets at multiple scales (10M, 100M, 1B, 10B) plus seed concepts and an SFT variant are in the Sutra Pedagogical Datasets collection.
- 2 replies
Post
3292
Reverse Engineering a $500M Mystery: From HashHop to Memory-Augmented Language Models
I wrote a deep dive into how Magic AI's 100M token context window might work, starting from their HashHop benchmark and building up to MALM - a Memory-Augmented Language Model.
Key insight: treating each key as a single token enables perfect retrieval at unlimited context lengths.
The article covers:
- How HashHop works and why its perfect accuracy is suspicious
- Building a tokenized solver that achieves 100% accuracy
- Scaling to MALM for real code search tasks
- Why this approach could handle 100M+ tokens
Read the full article: https://huggingface.co/blog/codelion/reverse-engineering-magic-hashhop
Try the model: codelion/malm-165m
Code: https://github.com/codelion/hash-hop
I wrote a deep dive into how Magic AI's 100M token context window might work, starting from their HashHop benchmark and building up to MALM - a Memory-Augmented Language Model.
Key insight: treating each key as a single token enables perfect retrieval at unlimited context lengths.
The article covers:
- How HashHop works and why its perfect accuracy is suspicious
- Building a tokenized solver that achieves 100% accuracy
- Scaling to MALM for real code search tasks
- Why this approach could handle 100M+ tokens
Read the full article: https://huggingface.co/blog/codelion/reverse-engineering-magic-hashhop
Try the model: codelion/malm-165m
Code: https://github.com/codelion/hash-hop
- 1 reply
Post
6197
Introducing Dhara-70M: A diffusion language model that achieves 3.8x higher throughput than autoregressive models!
Key findings from our research on optimal architectures for small language models:
→ Depth beats width: 32 layers outperforms 12 layers at the same parameter count
→ Best-in-class factuality: 47.5% on TruthfulQA
→ 10x training efficiency using WSD (Warmup-Stable-Decay) conversion
→ Canon layers add only 0.13% parameters but improve reasoning
We trained on 1B tokens using the optimal 50-30-20 dataset mix (PDFs + filtered web + educational content), then converted to diffusion with just 100M additional tokens.
Blog: https://huggingface.co/blog/codelion/optimal-model-architecture
Model: codelion/dhara-70m
Key findings from our research on optimal architectures for small language models:
→ Depth beats width: 32 layers outperforms 12 layers at the same parameter count
→ Best-in-class factuality: 47.5% on TruthfulQA
→ 10x training efficiency using WSD (Warmup-Stable-Decay) conversion
→ Canon layers add only 0.13% parameters but improve reasoning
We trained on 1B tokens using the optimal 50-30-20 dataset mix (PDFs + filtered web + educational content), then converted to diffusion with just 100M additional tokens.
Blog: https://huggingface.co/blog/codelion/optimal-model-architecture
Model: codelion/dhara-70m
- 1 reply
Post
2447
Introducing PTS Visualizer - an interactive tool for exploring how language models reason!
Visualize pivotal tokens, thought anchors, and reasoning circuits. See which tokens and sentences significantly impact success probability, explore embedding clusters, and trace reasoning step-by-step.
Try it: codelion/pts-visualizer
Explore PTS datasets:
- Qwen3-0.6B: codelion/Qwen3-0.6B-pts
- DeepSeek-R1: codelion/DeepSeek-R1-Distill-Qwen-1.5B-pts
Or upload your own JSONL files!
GitHub: https://github.com/codelion/pts
Visualize pivotal tokens, thought anchors, and reasoning circuits. See which tokens and sentences significantly impact success probability, explore embedding clusters, and trace reasoning step-by-step.
Try it: codelion/pts-visualizer
Explore PTS datasets:
- Qwen3-0.6B: codelion/Qwen3-0.6B-pts
- DeepSeek-R1: codelion/DeepSeek-R1-Distill-Qwen-1.5B-pts
Or upload your own JSONL files!
GitHub: https://github.com/codelion/pts
Post
2633
Recently, Essential AI released a new 8B base model EssentialAI/rnj-1 they highlighted the importance of data mix for pretraning -
"In the long run, we expect our methods to automatically represent, transform, and blend data to optimize measurable abilities in pre-training. Our work on modeling data taxonomies led to new approaches for jointly clustering and mixing data distributions under data repetition penalties. Many improvements in our STEM abilities can be traced back to this. "
This resonates with the recent work we did around optimal dataset mixing for pretraining where we saw have the right mix can increase the efficiency of training -
https://huggingface.co/blog/codelion/optimal-dataset-mixing
"In the long run, we expect our methods to automatically represent, transform, and blend data to optimize measurable abilities in pre-training. Our work on modeling data taxonomies led to new approaches for jointly clustering and mixing data distributions under data repetition penalties. Many improvements in our STEM abilities can be traced back to this. "
This resonates with the recent work we did around optimal dataset mixing for pretraining where we saw have the right mix can increase the efficiency of training -
https://huggingface.co/blog/codelion/optimal-dataset-mixing
Post
2707
NotebookLM's infographics feature is amazing, it generates poster-type images from any text. Here is one I tried for my new HF article on ellora - https://huggingface.co/blog/codelion/ellora-lora-recipes
Post
2353
Perplexity released a dataset (BrowseSafe) and benchmark to catch and prevent malicious prompt-injection instructions in real-time.
We trained a prompt injection classifier on BrowseSafe using adaptive-classifier with ModernBERT-base embeddings.
74.9% F1 on detecting prompt injection in web content.
Model -> adaptive-classifier/browsesafe
Dataset -> perplexity-ai/browsesafe-bench
Repo -> https://github.com/codelion/adaptive-classifier
We trained a prompt injection classifier on BrowseSafe using adaptive-classifier with ModernBERT-base embeddings.
74.9% F1 on detecting prompt injection in web content.
Model -> adaptive-classifier/browsesafe
Dataset -> perplexity-ai/browsesafe-bench
Repo -> https://github.com/codelion/adaptive-classifier
- 1 reply
Post
1630
I just published Ellora - 6 production-ready LoRA recipes for enhancing LLMs with specific capabilities. Each recipe costs under $100 to run and includes complete training code, data generation, and evaluation.
The 6 Recipes:
Recipe 1: Accuracy Recovery - Recover 75% of quantization losses with self-distillation
Recipe 2: Reasoning LoRA - Add structured thinking with GRPO (0% to 60% adoption, 75% quality boost)
Recipe 3: Tool Calling - Real execution on actual codebases
Recipe 4: Context Extension - Scale from 32K to 2M tokens (61x increase)
Recipe 5: Secure Code Generation - 97% vulnerability reduction using automated Semgrep analysis
Recipe 6: Execution-Aware World Models - Teaching models runtime behavior
Why Recipes?
Ellora provides methodologies, not frameworks. Use them with your existing tools (PEFT, LoRAX, vLLM, Unsloth, HuggingFace). Each recipe uses self-supervised data generation (Magpie approach) - no expensive human labeling required.
All recipes include Jupyter notebooks you can run immediately with clear success metrics.
GitHub: https://github.com/codelion/ellora
Full Article: https://huggingface.co/blog/codelion/ellora-lora-recipes
Built something with these recipes? I'd love to see what you create!
The 6 Recipes:
Recipe 1: Accuracy Recovery - Recover 75% of quantization losses with self-distillation
Recipe 2: Reasoning LoRA - Add structured thinking with GRPO (0% to 60% adoption, 75% quality boost)
Recipe 3: Tool Calling - Real execution on actual codebases
Recipe 4: Context Extension - Scale from 32K to 2M tokens (61x increase)
Recipe 5: Secure Code Generation - 97% vulnerability reduction using automated Semgrep analysis
Recipe 6: Execution-Aware World Models - Teaching models runtime behavior
Why Recipes?
Ellora provides methodologies, not frameworks. Use them with your existing tools (PEFT, LoRAX, vLLM, Unsloth, HuggingFace). Each recipe uses self-supervised data generation (Magpie approach) - no expensive human labeling required.
All recipes include Jupyter notebooks you can run immediately with clear success metrics.
GitHub: https://github.com/codelion/ellora
Full Article: https://huggingface.co/blog/codelion/ellora-lora-recipes
Built something with these recipes? I'd love to see what you create!
Post
2023
Introducing OpenEvolve Prompt Optimizer - a Space that automatically evolves and optimizes your prompts using OpenEvolve!
This tool uses OpenEvolve to iteratively improve prompts by testing them on real datasets and evolving better versions. No more manual prompt engineering guesswork - let OpenEvolve find the optimal prompts for you.
How it works:
- Enter your initial prompt using {input} as a placeholder for dataset inputs
- Input any HuggingFace dataset name you want to use for optimization
- Specify the dataset split and field names for your use case
- Click Optimize Prompt and the system will validate everything first
- Compare your initial prompt vs the evolved best prompt side-by-side
Try it here: algorithmicsuperintelligence/prompt-optimizer
OpenEvolve GitHub: https://github.com/algorithmicsuperintelligence/openevolve
This tool uses OpenEvolve to iteratively improve prompts by testing them on real datasets and evolving better versions. No more manual prompt engineering guesswork - let OpenEvolve find the optimal prompts for you.
How it works:
- Enter your initial prompt using {input} as a placeholder for dataset inputs
- Input any HuggingFace dataset name you want to use for optimization
- Specify the dataset split and field names for your use case
- Click Optimize Prompt and the system will validate everything first
- Compare your initial prompt vs the evolved best prompt side-by-side
Try it here: algorithmicsuperintelligence/prompt-optimizer
OpenEvolve GitHub: https://github.com/algorithmicsuperintelligence/openevolve
Post
2492
🎯 Introducing Chayan: A Calibrated 4-Model LLM Router Achieving 69% Accuracy on RouterArena
We're excited to share Chayan, a cost-efficient LLM router that intelligently routes queries between 4 models to maximize accuracy while minimizing cost. Chayan just submitted to the RouterArena leaderboard and achieved 69.05% accuracy on the benchmark!
🔗 Model: adaptive-classifier/chayan
🔗 Dataset: RouteWorks/RouterArena
📊 Performance Highlights
Chayan achieves impressive results on the RouterArena benchmark:
• 69.05% accuracy (would rank #1 on current leaderboard)
• $0.333 per 1K queries
• +12.07pp improvement over all-mini baseline (56.98%)
• 99% of perfect 2-model oracle performance at 57% lower cost
Compared to our previous 2-model router (61.43% accuracy), Chayan delivers +7.62pp improvement through smarter 4-model routing.
🧠 How It Works
Chayan uses an Adaptive K-NN classifier with prototype memory to route between 4 models:
• openai/gpt-4o-mini (fast & cheap)
• google/gemini-2.5-flash-lite (balanced)
• google/gemini-2.5-flash (capable)
• openai/gpt-4o (most powerful)
🚀 Getting Started
You can use Chayan directly from HuggingFace:
from adaptive_classifier import AdaptiveClassifier
Load Chayan
router = AdaptiveClassifier.load("adaptive-classifier/chayan")
Route a query
query = "What is the capital of France?"
predictions = router.predict(query, k=4)
Get top model recommendation
best_model = predictions[0][0]
print(f"Recommended model: {best_model}")
Built with the adaptive-classifier library: https://github.com/codelion/adaptive-classifier
We're excited to share Chayan, a cost-efficient LLM router that intelligently routes queries between 4 models to maximize accuracy while minimizing cost. Chayan just submitted to the RouterArena leaderboard and achieved 69.05% accuracy on the benchmark!
🔗 Model: adaptive-classifier/chayan
🔗 Dataset: RouteWorks/RouterArena
📊 Performance Highlights
Chayan achieves impressive results on the RouterArena benchmark:
• 69.05% accuracy (would rank #1 on current leaderboard)
• $0.333 per 1K queries
• +12.07pp improvement over all-mini baseline (56.98%)
• 99% of perfect 2-model oracle performance at 57% lower cost
Compared to our previous 2-model router (61.43% accuracy), Chayan delivers +7.62pp improvement through smarter 4-model routing.
🧠 How It Works
Chayan uses an Adaptive K-NN classifier with prototype memory to route between 4 models:
• openai/gpt-4o-mini (fast & cheap)
• google/gemini-2.5-flash-lite (balanced)
• google/gemini-2.5-flash (capable)
• openai/gpt-4o (most powerful)
🚀 Getting Started
You can use Chayan directly from HuggingFace:
from adaptive_classifier import AdaptiveClassifier
Load Chayan
router = AdaptiveClassifier.load("adaptive-classifier/chayan")
Route a query
query = "What is the capital of France?"
predictions = router.predict(query, k=4)
Get top model recommendation
best_model = predictions[0][0]
print(f"Recommended model: {best_model}")
Built with the adaptive-classifier library: https://github.com/codelion/adaptive-classifier
Post
4019
Want to experiment with pre-training dataset mixtures but don't want to process terabytes of data? We've got you covered.
We're releasing a collection of several carefully curated 1B token dataset samples specifically designed for rapid prototyping and pretraining experiments: https://huggingface.co/collections/codelion/pre-training-dataset-samples
These samples were created using reservoir sampling - an algorithm that guarantees statistically unbiased random samples from massive source datasets. This means results you get at the 1B token scale are representative of how these datasets behave at 100B+ token scales, letting you iterate quickly without the computational overhead.
The collection includes:
- finePDFs-1B: High-quality textbook-style educational content
- DCLM-baseline-1B: Filtered, diverse web content
- FineWeb-Edu-1B: Curated educational web resources
We used these exact samples to run 50+ systematic experiments on dataset mixing strategies, ultimately discovering that a 50-30-20 mixture of finePDFs + DCLM-baseline + FineWeb-Edu achieves 90%+ of GPT-2's performance with just 1/10th the training data.
Whether you're researching optimal data mixtures, testing curriculum learning strategies, or just want to quickly prototype a pretraining run, these samples give you a solid foundation to start experimenting immediately.
Read the full story of how we used these datasets to find the optimal pretraining recipe: https://huggingface.co/blog/codelion/optimal-dataset-mixing
We're releasing a collection of several carefully curated 1B token dataset samples specifically designed for rapid prototyping and pretraining experiments: https://huggingface.co/collections/codelion/pre-training-dataset-samples
These samples were created using reservoir sampling - an algorithm that guarantees statistically unbiased random samples from massive source datasets. This means results you get at the 1B token scale are representative of how these datasets behave at 100B+ token scales, letting you iterate quickly without the computational overhead.
The collection includes:
- finePDFs-1B: High-quality textbook-style educational content
- DCLM-baseline-1B: Filtered, diverse web content
- FineWeb-Edu-1B: Curated educational web resources
We used these exact samples to run 50+ systematic experiments on dataset mixing strategies, ultimately discovering that a 50-30-20 mixture of finePDFs + DCLM-baseline + FineWeb-Edu achieves 90%+ of GPT-2's performance with just 1/10th the training data.
Whether you're researching optimal data mixtures, testing curriculum learning strategies, or just want to quickly prototype a pretraining run, these samples give you a solid foundation to start experimenting immediately.
Read the full story of how we used these datasets to find the optimal pretraining recipe: https://huggingface.co/blog/codelion/optimal-dataset-mixing
- 1 reply
Post
293
MARS Achieves Strong Results on Google DeepMind's IMO-Bench
We evaluated OptiLLM's MARS (Multi-Agent Reasoning System) approach on IMO-Bench, Google DeepMind's challenging mathematical reasoning benchmark with International Mathematical Olympiad-level problems.
What is MARS?
MARS is a multi-agent reasoning technique that works with any LLM. It uses 3 parallel reasoning agents that independently solve problems, then verifies their solutions through consensus and iterative refinement. The key advantage: it's model-agnostic and can be applied to any base model through OptiLLM's inference proxy.
Results on IMO-Bench:
AnswerBench (400 short-answer problems):
MARS: 36.0% (144/400 correct)
Baseline: 24.5% (98/400 correct)
Improvement: +11.5pp across all domains
Category breakdown:
- Algebra: 33% (vs 21% baseline)
- Combinatorics: 26% (vs 19% baseline)
- Geometry: 43% (vs 28% baseline)
- Number Theory: 42% (vs 30% baseline)
ProofBench (60 proof construction problems):
MARS: 26.7% (16/60 correct)
Baseline: 18.3% (11/60 correct)
Improvement: +8.4pp
Category breakdown:
- Number Theory: 42.9% (vs 14.3% baseline)
- Combinatorics: 37.5% (vs 31.2% baseline)
- Algebra: 18.8% (vs 25.0% baseline)
- Geometry: 7.1% (vs 0.0% baseline)
All results achieved using google/gemini-2.5-flash-lite-preview-09-2025 as the base model. The same MARS approach can enhance reasoning for any model through OptiLLM's OpenAI-compatible API.
Datasets available at:
AnswerBench: huggingface.co/datasets/Hwilner/imo-answerbench
ProofBench: huggingface.co/datasets/Hwilner/imo-proofbench
Try it yourself:
python optillm.py --approach mars --model google/gemini-2.5-flash-lite-preview-09-2025
Or via API with approach prefix:
model: "mars-google/gemini-2.5-flash-lite-preview-09-2025"
Full evaluation code and results available at: github.com/algorithmicsuperintelligence/optillm
We evaluated OptiLLM's MARS (Multi-Agent Reasoning System) approach on IMO-Bench, Google DeepMind's challenging mathematical reasoning benchmark with International Mathematical Olympiad-level problems.
What is MARS?
MARS is a multi-agent reasoning technique that works with any LLM. It uses 3 parallel reasoning agents that independently solve problems, then verifies their solutions through consensus and iterative refinement. The key advantage: it's model-agnostic and can be applied to any base model through OptiLLM's inference proxy.
Results on IMO-Bench:
AnswerBench (400 short-answer problems):
MARS: 36.0% (144/400 correct)
Baseline: 24.5% (98/400 correct)
Improvement: +11.5pp across all domains
Category breakdown:
- Algebra: 33% (vs 21% baseline)
- Combinatorics: 26% (vs 19% baseline)
- Geometry: 43% (vs 28% baseline)
- Number Theory: 42% (vs 30% baseline)
ProofBench (60 proof construction problems):
MARS: 26.7% (16/60 correct)
Baseline: 18.3% (11/60 correct)
Improvement: +8.4pp
Category breakdown:
- Number Theory: 42.9% (vs 14.3% baseline)
- Combinatorics: 37.5% (vs 31.2% baseline)
- Algebra: 18.8% (vs 25.0% baseline)
- Geometry: 7.1% (vs 0.0% baseline)
All results achieved using google/gemini-2.5-flash-lite-preview-09-2025 as the base model. The same MARS approach can enhance reasoning for any model through OptiLLM's OpenAI-compatible API.
Datasets available at:
AnswerBench: huggingface.co/datasets/Hwilner/imo-answerbench
ProofBench: huggingface.co/datasets/Hwilner/imo-proofbench
Try it yourself:
python optillm.py --approach mars --model google/gemini-2.5-flash-lite-preview-09-2025
Or via API with approach prefix:
model: "mars-google/gemini-2.5-flash-lite-preview-09-2025"
Full evaluation code and results available at: github.com/algorithmicsuperintelligence/optillm
Post
3626
On this day in 2019, OpenAI released the final GPT-2 model as part of their staged release. I still remember that November well - so much was happening, but GPT-2's release felt like a watershed moment for the field. It showed us what was possible with carefully trained language models.
To recreate some of that GPT-2 magic, I recently tackled an interesting challenge: can you pretrain a language model with just 1 billion tokens - roughly 1/10th of what GPT-2 used - and still get comparable performance? After 50+ systematic experiments testing different dataset mixtures, the answer is yes.
The result is codelion/gpt-2-70m, which achieves over 90% of GPT-2's benchmark performance despite being trained on 10x less data. The key was finding the optimal dataset composition: 50% high-quality textbook PDFs, 30% filtered web content, and 20% educational resources. It even beats GPT-2 on TruthfulQA (47.31% vs 40.69%).
If you're interested in the full story of how we discovered this optimal mixture and why curriculum learning catastrophically failed, check out the complete article: https://huggingface.co/blog/codelion/optimal-dataset-mixing
Sometimes less really is more - when you mix it right.
To recreate some of that GPT-2 magic, I recently tackled an interesting challenge: can you pretrain a language model with just 1 billion tokens - roughly 1/10th of what GPT-2 used - and still get comparable performance? After 50+ systematic experiments testing different dataset mixtures, the answer is yes.
The result is codelion/gpt-2-70m, which achieves over 90% of GPT-2's benchmark performance despite being trained on 10x less data. The key was finding the optimal dataset composition: 50% high-quality textbook PDFs, 30% filtered web content, and 20% educational resources. It even beats GPT-2 on TruthfulQA (47.31% vs 40.69%).
If you're interested in the full story of how we discovered this optimal mixture and why curriculum learning catastrophically failed, check out the complete article: https://huggingface.co/blog/codelion/optimal-dataset-mixing
Sometimes less really is more - when you mix it right.
- 1 reply
Post
402
The 1 Billion Token Challenge: Finding the Perfect Pre-training Mix
We trained a GPT-2 model to 90%+ performance using just 1/10th the training data through 50+ systematic experiments on dataset mixing strategies.
Key Finding:
A static mix of 50% finePDFs + 30% DCLM-baseline + 20% FineWeb-Edu consistently outperforms complex curriculum learning approaches. Static mixing is simpler, faster, and avoids catastrophic failures from hard distribution shifts.
Results:
Our GPT-2-70M model (70M parameters, 1B tokens) scores 38.15% on benchmarks vs GPT-2's 39.13% - only 0.98 points behind despite 10x less data and 44% fewer parameters. It even beats GPT-2 on TruthfulQA (47.31% vs 40.69%).
The takeaway: careful dataset curation matters more than total data volume.
Model: codelion/gpt-2-70m
Datasets: https://huggingface.co/collections/codelion/pre-training-dataset-samples
Full blog: https://huggingface.co/blog/codelion/optimal-dataset-mixing
We trained a GPT-2 model to 90%+ performance using just 1/10th the training data through 50+ systematic experiments on dataset mixing strategies.
Key Finding:
A static mix of 50% finePDFs + 30% DCLM-baseline + 20% FineWeb-Edu consistently outperforms complex curriculum learning approaches. Static mixing is simpler, faster, and avoids catastrophic failures from hard distribution shifts.
Results:
Our GPT-2-70M model (70M parameters, 1B tokens) scores 38.15% on benchmarks vs GPT-2's 39.13% - only 0.98 points behind despite 10x less data and 44% fewer parameters. It even beats GPT-2 on TruthfulQA (47.31% vs 40.69%).
The takeaway: careful dataset curation matters more than total data volume.
Model: codelion/gpt-2-70m
Datasets: https://huggingface.co/collections/codelion/pre-training-dataset-samples
Full blog: https://huggingface.co/blog/codelion/optimal-dataset-mixing