| --- |
| pretty_name: res-vllm |
| --- |
| |
| # ryzax/res-vllm |
|
|
| Copy of [`ryzax/res`](https://huggingface.co/datasets/ryzax/res) with vLLM-tool AIME25 runs swapped into the SUMM and LASTK folders. Original `ryzax/res` is unchanged. |
|
|
| All swapped/added runs are Qwen3-1.7B, AIME25 (30 problems × 64 samples), window 4096, max gen 262144, `BUDGET_FORCE_SAVING=256` for summary variants. |
|
|
| ## Swaps |
|
|
| | Path | What it is now | AIME25 avg@64 | cov@64 | maj@64 | |
| |---|---|---:|---:|---:| |
| | `Qwen3-1.7B-SUMM/w4096s256` | force256 + `SAVING_PROMPT=context_system` (evaluate-context example on **last save only**) | 26.5% | 56.7% | 40.0% | |
| | `Qwen3-1.7B-LASTK/w4096k256` | last-k 256 with subtract | 23.8% | 50.0% | 36.7% | |
|
|
| Previously these folders held the Hugging Face Transformers SUMM (`w4096s256`) and LASTK (`w4096k256`) runs from Oct 2025. |
|
|
| ## SUMM ablations |
|
|
| | Path | Prompt | avg@64 | cov@64 | maj@64 | |
| |---|---|---:|---:|---:| |
| | `Qwen3-1.7B-SUMM/w4096s256context` | force256 + `context` | 26.4% | 53.3% | 36.7% | |
| | `Qwen3-1.7B-SUMM/w4096s256context_maybe` | force256 + `context_maybe` | 25.8% | 46.7% | 40.0% | |
| | `Qwen3-1.7B-SUMM/w4096s256contextnoshot` | force256 + `context`, no few-shot example in the system prompt | 23.2% | 33.3% | 33.3% | |
| | `Qwen3-1.7B-SUMM/w4096s256context_system` | force256 + `context_system`, evaluate-context example on **every save** | 22.8% | 36.7% | 33.3% | |
|
|
| ## LASTK ablations |
|
|
| | Path | Setting | avg@64 | cov@64 | maj@64 | |
| |---|---|---:|---:|---:| |
| | `Qwen3-1.7B-LASTK/w4096k256nosubtract` | last-k 256, no subtract | 26.3% | 53.3% | 43.3% | |
|
|
| Dense / sliding-window folders are the same files as `ryzax/res`. |
|
|
| ## Length metrics and budget-clipped avg |
|
|
| `results_*.json` scored only the last assistant window (~3k tokens), so `tok@` is too small and `avg@` jumps by 4k. Use: |
|
|
| - `results_*.tokfixed.json` — full-trace `tok@` / `too_long@` from `tool_usage.generated_tokens` (accuracy unchanged) |
| - `results_concatted_tok262144.json` — `aime/utils.py` rerun on concatenated saving windows + final (the old `completions_to_return_concatted` path). Full-budget avg matches the original; mid-budget avg rises with true length. |
|
|
| | Path | tok@262144@64 | avg@4k@64 | avg@32k@64 | avg@262k@64 | |
| |---|---:|---:|---:|---:| |
| | `Qwen3-1.7B-SUMM/w4096s256` | 123615 | 14.8% | 22.3% | 26.5% | |
| | `Qwen3-1.7B-SUMM/w4096s256context` | 122880 | 15.5% | 22.4% | 26.4% | |
| | `Qwen3-1.7B-SUMM/w4096s256context_maybe` | 139369 | 15.6% | 22.9% | 25.8% | |
| | `Qwen3-1.7B-SUMM/w4096s256contextnoshot` | 153125 | 4.7% | 19.2% | 23.2% | |
| | `Qwen3-1.7B-LASTK/w4096k256` | 129723 | 3.5% | 20.1% | 23.9% | |
| | `Qwen3-1.7B-LASTK/w4096k256nosubtract` | 128707 | 3.2% | 20.4% | 26.3% | |
| | `Qwen3-1.7B-SUMM/w4096s256context_system` | 119720 | 15.7% | 20.8% | 22.8% | |
|
|
| `w4096s256context_system` scored concatenated windows live (`results_*.json` already has full-trace `tok@` and rising mid-budget avg), so it has no `tokfixed` / `results_concatted` sidecars. Applying the evaluate-context example on every save hurts vs last-save-only `w4096s256` (22.8% vs 26.5% avg@64). |
|
|
