| --- |
| license: mit |
| task_categories: |
| - text-generation |
| language: |
| - en |
| tags: |
| - math |
| - reasoning |
| - sft |
| - preference-pairs |
| pretty_name: "Axiom-World Phase-1 general reasoning mixture" |
| --- |
| |
| # axiom-general-posttrain — Phase-1 general reasoning mixture (protocol v1) |
|
|
| The Phase-1 ("general warm-start") training mixture of the Axiom-World |
| project: a pre-registered, weighted blend of public math-reasoning corpora |
| with **dataset-derived gold answers only** (no LLM-generated labels). |
|
|
| - **Code:** https://github.com/m97j/axiom-world (tag `v1.0.0`) |
|
|
| ## Contents |
|
|
| | path | what | |
| |---|---| |
| | `v1/` | SFT mixture — GSM8K (weight 0.6) + MATH-algebra (weight 0.4), target 8,000 records, prompt-level dedup; 500-record held-out retention suite | |
| | `pref-v1/` | rejection-sampling preference pairs mined on the same distribution (used by the B2/B3 arms) | |
|
|
| Records are chat-rendered (Qwen3 template) with answers normalized to the |
| source datasets' gold format. Fingerprints are recorded in each consuming |
| run's `lineage.json`. |
|
|
| ## Source datasets & licenses |
|
|
| - **GSM8K** (`openai/gsm8k`) — MIT license. Cobbe et al., 2021. |
| - **MATH**, algebra subject (`EleutherAI/hendrycks_math`) — MIT license. |
| Hendrycks et al., 2021. |
|
|
| This repository redistributes preprocessed derivatives of the above under |
| their original MIT terms; all credit for the underlying problems goes to the |
| original authors. |
|
|
| ## Citation |
|
|
| Cite the original datasets (Cobbe et al. 2021; Hendrycks et al. 2021) and the |
| Axiom-World tech report. |
|
|
