sci-eval-final / README.md
JLiangHe's picture
Organize four final Sci-LLM stages; select 12x24 step35 agentic RL and preserve intermediates
f88b1bd verified
|
Raw History Blame Contribute Delete
1.63 kB
# Science evaluation results
## Final Sci-LLM checkpoints
| Stage | Selected checkpoint | Research rubric | Olympiad | arXiv Math | HLE | SciCoding-Agent | TB 2.1 |
|---|---|---:|---:|---:|---:|---:|---:|
| [sft-mix](sft-mix/) | Mix SFT | 29.50% | 61.17% | 26.46% | 30.69% | 20.00% | 22.47% |
| [sft-agentic](sft-agentic/) | Agentic SFT — step 1,144 | 29.59% | 61.40% | 25.87% | 31.25% | 37.00% | 27.53% |
| [rl-reasoning](rl-reasoning/) | Reasoning RL — step 180 | 35.70% | 63.43% | 31.07% | 33.88% | 30.00% | 21.91% |
| [rl-agentic](rl-agentic/) | Agentic RL 12×24 — step 35 | 34.23% | 62.93% | 30.29% | 31.19% | 46.00% | 28.09% |
Each stage has its final benchmark results directly under the stage folder and an `intermediate/` directory for previously published intermediate checkpoints. Results use the full planned denominator. Research is mean rubric, not threshold success rate; FS uses 30 trials, arXiv Math 20, HLE one, and agentic benchmarks two.
## Comparison models
- [Gemma-4-12B-it](Gemma-4-12B-it/)
- [NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16](NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16/)
- [Nemotron-Cascade-2-30B-A3B](Nemotron-Cascade-2-30B-A3B/)
- [Olmo-3-32B-Think](Olmo-3-32B-Think/)
- [Qwen3.5-35B-A3B](Qwen3.5-35B-A3B/)
- [Qwen3.5-9B](Qwen3.5-9B/)
- [gpt-oss-20b](gpt-oss-20b/)
- [granite-4.2-8b](granite-4.2-8b/)
## Protocol and archives
See [agentic configuration alignment](AGENTIC_CONFIG_ALIGNMENT.md) and each benchmark’s frozen configs. Prior alternative runs and excluded comparison models are in [backup](backup/). Historical results remain recoverable from repository revisions.