61 lines
3.2 KiB
Markdown
61 lines
3.2 KiB
Markdown
---
|
|
license: other
|
|
license_name: lfm1.0
|
|
license_link: https://www.liquid.ai/legal/lfm-license
|
|
base_model:
|
|
- anakin87/LFM2-2.6B-ttt-rl-merged
|
|
library_name: transformers
|
|
tags:
|
|
- rl
|
|
- cispo
|
|
- tictactoe
|
|
- rlvr
|
|
pipeline_tag: text-generation
|
|
language:
|
|
- en
|
|
---
|
|
|
|
# LFM2-2.6B-mr-tictactoe
|
|
|
|
A 2.6B parameter model that plays near-perfect Tic Tac Toe, outperforming `openai/gpt-5-mini` on this task.
|
|
|
|
Built from [LiquidAI/LFM2-2.6B](https://huggingface.co/LiquidAI/LFM2-2.6B) through a full training pipeline: Supervised Fine-Tuning on synthetic data, followed by two rounds of Reinforcement Learning (CISPO) in a verifiable Tic Tac Toe environment.
|
|
|
|
This model was developed as part of 🎓 **[LLM RL Environments Lil Course](https://github.com/anakin87/llm-rl-environments-lil-course)**, a hands-on course on building RL environments for Language Models, where models learn from rewards, not examples. It walks through the full process of turning a small open model into a specialist that outperforms a large proprietary one on a specific task (Tic Tac Toe).
|
|
|
|
🤗🕹️ **[Play against Mr. Tic Tac Toe](https://huggingface.co/spaces/anakin87/LFM2-2.6B-mr-tictactoe)**
|
|
|
|
|
|
|
|
## Training pipeline
|
|
|
|
| Step | Model | Method |
|
|
|------|-------|--------|
|
|
| 1. SFT warm-up | [anakin87/LFM2-2.6B-ttt-sft](https://huggingface.co/anakin87/LFM2-2.6B-ttt-sft) | SFT on 174 synthetic games from `gpt-5-mini` |
|
|
| 2. RL round 1 | [anakin87/LFM2-2.6B-ttt-rl](https://huggingface.co/anakin87/LFM2-2.6B-ttt-rl) + [merged](https://huggingface.co/anakin87/LFM2-2.6B-ttt-rl-merged) | CISPO, 600 steps, opponents at 20-70% random |
|
|
| 3. RL round 2 | [anakin87/LFM2-2.6B-ttt-rl-2](https://huggingface.co/anakin87/LFM2-2.6B-ttt-rl-2) + this model | CISPO, 400 steps, opponents at 0-25% random, temp 1.25 |
|
|
|
|
## Evaluation
|
|
|
|
100 games per setting. The model plays as X (first mover) against a Minimax-based opponent.
|
|
|
|
| **Model vs random opponent** | **% Wins** | **% Draws** | **% Losses** | **% Follows format** | **% Games w invalid moves** |
|
|
|------------------------------|------------|-------------|--------------|----------------------|---------------------|
|
|
| openai/gpt-5-mini | 90 | 9 | 1 | 100 | 0 |
|
|
| LiquidAI/LFM2-2.6B | 40 | 11 | 49 | 27.8 | 40 |
|
|
| **anakin87/LFM2-2.6B-mr-tictactoe** | **90** | **10** | **0** | **100** | **0** |
|
|
| | | | | | |
|
|
| **Model vs optimal opponent** | **% Wins** | **% Draws** | **% Losses** | **% Follows format** | **% Games w invalid moves** |
|
|
| openai/gpt-5-mini | 0 | 76 | 24 | 100 | 0 |
|
|
| LiquidAI/LFM2-2.6B | 0 | 11 | 89 | 24.7 | 43 |
|
|
| **anakin87/LFM2-2.6B-mr-tictactoe** | **0** | **97** | **3** | **99.8** | **0** |
|
|
|
|
## Training details
|
|
|
|
- **Algorithm:** CISPO (two rounds), using [Verifiers](https://github.com/PrimeIntellect-ai/verifiers) RLTrainer
|
|
- **Environment:** [anakin87/tictactoe](https://app.primeintellect.ai/dashboard/environments/anakin87/tictactoe) (Verifiers environment)
|
|
- **LoRA rank:** 8
|
|
- **Hardware:** 2x NVIDIA RTX Pro 6000 (round 1), 2x NVIDIA H200 (round 2)
|
|
- **Training time:** ~8 hours per round
|
|
- **W&B project:** [LFM2-2.6B Tic Tac Toe](https://wandb.ai/stefanofiorucci/LFM2-2.6B%20Tic%20Tac%20Toe/table)
|