id string | domain string | generator string | upstream_source string | expected_answer string | n_turns int32 | messages list |
|---|---|---|---|---|---|---|
345de401-3451-542f-a4d9-bf6a65f9d861 | verification | DeepSeek-V4-Pro | AoPS | 1 | [
{
"role": "system",
"content": "You are an expert mathematical assistant. Provide rigorous, complete proofs.",
"reasoning_content": ""
},
{
"role": "user",
"content": "## Instruction\n\nYour task is to evaluate the quality of a solution to a problem. The problem may ask for a proof of statem... | |
45f3f047-6b7c-5f4c-9e97-d3bcd992f0b9 | meta_verification | DeepSeek-V4-Pro | AoPS | 1 | [
{
"role": "system",
"content": "You are an expert mathematical assistant. Provide rigorous, complete proofs.",
"reasoning_content": ""
},
{
"role": "user",
"content": "You are given a \"problem\", \"solution\", and \"solution evaluation\", and you need to assess the whether this \"solution e... | |
94166e73-7f8f-5ff4-bb3b-1fdd3675f4c1 | verification | DeepSeek-V4-Pro | AoPS | 1 | [{"role":"system","content":"You are an expert mathematical assistant. Provide rigorous, complete pr(...TRUNCATED) | |
ultra_dsv4_cot_aops_rs3_000012921 | aops_cot | DeepSeek-V4-Pro | AoPS | k=2 | 1 | [{"role":"system","content":"You are an expert mathematical assistant. Provide rigorous, complete so(...TRUNCATED) |
bd280c98-0d5b-5ab0-be75-71a8aa91af4a | verification | DeepSeek-V4-Pro | AoPS | 1 | [{"role":"system","content":"You are an expert mathematical assistant. Provide rigorous, complete pr(...TRUNCATED) | |
95cba500-8248-52b4-a0cc-98fce2ed3d86 | verification | DeepSeek-V4-Pro | AoPS | 1 | [{"role":"system","content":"You are an expert mathematical assistant. Provide rigorous, complete pr(...TRUNCATED) | |
ultra_dsv4_cot_aops_rs3_000009563 | aops_cot | DeepSeek-V4-Pro | AoPS | 125^{\circ} | 1 | [{"role":"system","content":"You are an expert mathematical assistant. Provide rigorous, complete so(...TRUNCATED) |
60fdaa46-657e-53dd-b484-83f17bc391c3 | proof | DeepSeek-V4-Pro | AoPS | 1 | [{"role":"system","content":"You are an expert mathematical assistant. Provide rigorous, complete pr(...TRUNCATED) | |
b1c2d699-2b9f-5eb3-9db9-2e6c584f3602 | proof | DeepSeek-V4-Pro | AoPS | 1 | [{"role":"system","content":"You are an expert mathematical assistant. Provide rigorous, complete pr(...TRUNCATED) | |
3ff1db68-1002-50df-843c-02dd144cf9de | meta_verification | DeepSeek-V4-Pro | AoPS | 1 | [{"role":"system","content":"You are an expert mathematical assistant. Provide rigorous, complete pr(...TRUNCATED) |
yccchen-stage2-sft
The stage-2 (256K long-context) SFT training mix for opd-32b (Olmo-3 32B backbone with a transplanted DeepSeek-V4 tokenizer), in the model's native DeepSeek-R1 chat format. This is the materialized, globally-shuffled training set — repeats are baked in and rows are randomly ordered.
- 159,394 conversations · ~8.6 B tokens (opd-32b tokenizer).
- Assistant reasoning in a separate
reasoning_contentfield (<think>…</think>), answer incontent. - Per-source system prompt; task instruction native in the user turn. No tools.
from datasets import load_dataset
ds = load_dataset("chankhavu/yccchen-stage2-sft", split="train")
Mixture (materialized)
| source | domain | rows | system prompt |
|---|---|---|---|
| Nemotron-Math-Proofs-v2 | proof (×2 upweight) |
49,392 | "…complete proofs." |
| Nemotron-Math-Proofs-v2 | verification |
28,865 | "…complete proofs." |
| Nemotron-Math-Proofs-v2 | meta_verification |
29,176 | "…complete proofs." |
| Nemotron-SFT-Math-v4 | aops_cot |
51,961 | "…complete solutions." |
| Total | 159,394 |
proofis upweighted ×2 (materialized as duplicate rows) — the long-context proofs are the focus of the 256K stage. All rows are globally shuffled (scatter + per-shard shuffle), so the duplicates are spread and sources are interleaved on-disk.- Proofs use
"You are an expert mathematical assistant. Provide rigorous, complete proofs."; AoPS CoT (numeric-answer solutions) use the"…complete solutions."variant.
Chat format (opd-32b / DeepSeek-R1)
<|begin▁of▁sentence|>{system}<|User|>{problem}<|Assistant|><think>{reasoning}</think>{answer}<|end▁of▁sentence|>
— <think>/</think> are dedicated tokens (bos=0, eos=1). Supervised loss on the assistant span only
({reasoning}</think>{answer}<|end▁of▁sentence|>); everything up to <|Assistant|><think> is masked.
Pre-tokenized olmo-core companion: chankhavu/yccchen-stage2-olmocore-256k (docs >262144 tokens dropped).
Schema: id, domain, generator, upstream_source, expected_answer, n_turns, messages.
CC-BY-4.0 (inherits upstream Nemotron licenses).
- Downloads last month
- 5