Card: add 'Models used' section
Browse filesNames the generator, judge and tokenizer models (with links) for each pipeline stage.
README.md
CHANGED
|
@@ -295,6 +295,16 @@ text = f"Problem: {r['rewritten_problem']}\n\nSolution: {r['rewritten_solution']
|
|
| 295 |
bad = d.filter(lambda r: r["judge_correctness"] == "no")
|
| 296 |
```
|
| 297 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 298 |
## How it was built
|
| 299 |
|
| 300 |
**Stage 1 — seed generation.** 60,000,000 problems from 122 named generators across
|
|
|
|
| 295 |
bad = d.filter(lambda r: r["judge_correctness"] == "no")
|
| 296 |
```
|
| 297 |
|
| 298 |
+
## Models used
|
| 299 |
+
|
| 300 |
+
| Role | Model | Settings |
|
| 301 |
+
|---|---|---|
|
| 302 |
+
| Rewrite (stage 2) | [`RedHatAI/gemma-4-31B-it-FP8-dynamic`](https://huggingface.co/RedHatAI/gemma-4-31B-it-FP8-dynamic) (FP8 quantization of [`google/gemma-4-31B-it`](https://huggingface.co/google/gemma-4-31B-it)) | temperature 0.8, served with vLLM |
|
| 303 |
+
| Judge (stage 3) | [`RedHatAI/gemma-4-31B-it-FP8-dynamic`](https://huggingface.co/RedHatAI/gemma-4-31B-it-FP8-dynamic) (FP8 quantization of [`google/gemma-4-31B-it`](https://huggingface.co/google/gemma-4-31B-it)), the same model as the rewriter | temperature 0.1, strict-JSON rubric |
|
| 304 |
+
| Token counts (`token_count`) | Qwen2.5 tokenizer ([`Qwen/Qwen2.5-7B-Instruct`](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct)) | |
|
| 305 |
+
|
| 306 |
+
Seed problems (stage 1) come from deterministic code, not from a model.
|
| 307 |
+
|
| 308 |
## How it was built
|
| 309 |
|
| 310 |
**Stage 1 — seed generation.** 60,000,000 problems from 122 named generators across
|