---
license: apache-2.0
base_model: Qwen/Qwen2.5-Coder-3B-Instruct
library_name: peft
tags:
  - competitive-programming
  - codeforces
  - qlora
datasets:
  - open-r1/codeforces-cots
---

# Ladder-3B (unevaluated)

A QLoRA adapter for `Qwen2.5-Coder-3B-Instruct`, trained on verified Codeforces
reasoning traces with [Ladder](https://github.com/NiLabs-Models/ladder).

> **This model has not been evaluated.** There is no pass@1 for it, and none is
> claimed. It is stored here so the weights are not stranded on a temporary
> disk, and so any GPU can pick it up and score it. The repository stays private
> until there are measured numbers to publish with it.

## Training

| | |
| --- | --- |
| Base | `unsloth/Qwen2.5-Coder-3B-Instruct-bnb-4bit` |
| Method | QLoRA, 4-bit, rank 32, loss on the assistant turn only |
| Data | 2,271 traces, each verified by executing its solution against the problem's own tests |
| Steps | 150, effective batch 16, context 8,192 |
| Hardware | Tesla T4, 5.05 h, 594.9 tok/s, peak VRAM 6.78 GB |

## Evaluating it

```bash
git clone https://github.com/NiLabs-Models/ladder
ladder eval --config configs/ladder-3b-kaggle.yaml                  # base
ladder eval --config configs/ladder-3b-kaggle.yaml --adapter <this> # tuned
```

Scoring runs generated programs against real Codeforces test cases. No
model-as-judge, no similarity metric.

## Known limitations

- Python only; most competitive programming is C++.
- Trained on the 30% of the corpus short enough to fit 8,192 tokens. The median
  trace is ~13,770 tokens, so the hardest problems are disproportionately absent.
- Codeforces `checker` and interactive problems are excluded from evaluation.