Danikm/minimax-m3-routing

MiniMax-M3 router inputs — 10M generated tokens, one MoE layer per token

For 10,104,532 tokens generated by MiniMax-M3 (bf16) on GPQA, HumanEval, general chat prompts and τ³-bench agent conversations, this dataset stores the input to the MoE router (the 6144-dim hidden state the gate sees) for one MoE layer per token, plus the top-4 experts that layer selected.

MiniMax-M3: 60 decoder layers, layers 3..59 are MoE (128 routed experts, top-4, sigmoid scoring with routing bias, 1 shared expert). Router input = output of post_attention_layernorm = argument of the MoE block; the gate computes fp32 logits x @ W_gate^T, then sigmoid(logits) + e_score_correction_bias → top-4.

Layer per token: token at position p of call r stores MoE layer moe_layers[(p + r) % 57], so every layer gets ~1/57 of the tokens, spread over all calls (177,126–177,367 tokens per layer).

source	calls	generated tokens
general	871	2,931,430
gpqa	287	2,751,040
humaneval	1,717	2,597,062
tau3 (airline)	1,028	257,502
tau3 (banking_knowledge)	2,274	881,738
tau3 (retail)	2,844	412,834
tau3 (telecom)	2,287	272,926
total	11,308	10,104,532

120 shards, 124 GB.

Files

shard_NNNNN.safetensors      ~80K rows each
    router_input           bf16  [N, 6144]  input to the gate of decoder layer `layer`
    layer                  uint8 [N]        decoder layer index (3..59)
    expert_ids             uint8 [N, 4]     top-4 experts vLLM selected (same pass)
    expert_ids_recomputed  uint8 [N, 4]     top-4 recomputed from router_input with that layer's gate weights
    expert_ids_generation  uint8 [N, 4]     top-4 recorded while the token was being generated
    record_id int32 [N]    call id (-> index.jsonl)
    position  int32 [N]    position in the call's token sequence
    token_id  int32 [N]    the generated token at that position (MiniMax-M3 tokenizer)
index.jsonl                  one line per call: record_id, tag (dataset), role, meta (item id, sample, ...),
                             prompt_len, completion_len, finish_reason, shard, rows, checks
layout.json                  MoE layer list and the layer-assignment rule

Only model-generated tokens are included (no prompt tokens). Row position p is the token at p fed to the model while producing token p+1; the final sampled token of a call has no row.

Loading

from huggingface_hub import hf_hub_download, snapshot_download
from safetensors.torch import load_file      # router_input is bfloat16 -> use the torch loader

t = load_file(hf_hub_download("Danikm/minimax-m3-routing", "shard_00000.safetensors", repo_type="dataset"))
x, layer = t["router_input"], t["layer"]     # [N, 6144] bf16, [N]
x30 = x[layer == 30]                         # router inputs of decoder layer 30

# everything (~124 GB):  snapshot_download("Danikm/minimax-m3-routing", repo_type="dataset", local_dir="m3_router_inputs")

How it was made

1. Generation. MiniMax-M3 bf16 on 8×H200, vLLM 0.23.1rc1 (TP8, block size 128, thinking enabled, temperature 1.0, top-p 0.95 as recommended on the model card).
   • GPQA main (448 questions, simple-evals multiple-choice prompt, choices reshuffled per sample), HumanEval (164 problems), general chat (AlpacaEval 805, Arena-Hard v0.1 500 + v2.0 750, MT-Bench 80 with both turns). Items visited round-robin with repeated samples; ≤32K new tokens per answer.
   • τ³-bench (tau2-bench harness): airline, retail, telecom, banking_knowledge (bm25 retrieval); MiniMax-M3 is the agent (thinking on) and also the user simulator (thinking off; user-simulator tokens are not in this dataset). A few banking conversations hit the 131K context limit and end there.
   • Every token's top-4 experts in all 57 MoE layers were captured with --enable-return-routed-experts.
2. Router inputs. vLLM cannot return hidden states during generation, so each call was replayed as a single teacher-forced prefill of its exact token sequence (vLLM eager mode, forward pre-hooks on every MoE block on TP rank 0, prefix cache reused within a conversation) and the router input of the assigned layer was saved for every generated position.

Validation

• expert_ids_recomputed (top-4 from the stored router_input with the checkpoint gate) equals vLLM's expert_ids on 100.000 % of rows (min per call 98.57 %); differences are fp32 near-ties.
• expert_ids vs expert_ids_generation: mean expert overlap 94.3 %. The replay is a different batch shape than generation, and bf16 nondeterminism flips near-tied experts (~3-4 % of expert slots, measured separately by re-prefilling identical sequences). router_input is exactly consistent with expert_ids.
• Positions and per-call row counts were checked against the generation-time routing record for every call.

Licenses and terms

Model outputs and activations of MiniMax-M3 — MiniMax Community License (non-commercial; commercial use requires a "Built with MiniMax M3" notice and notifying/authorization by MiniMax — see the license). Prompt sources keep their own terms: GPQA (CC BY 4.0; its authors ask that GPQA examples not be revealed in plain text online — this dataset contains model-generated tokens on GPQA prompts, not the prompts), HumanEval (MIT), AlpacaEval (CC BY-NC 4.0), Arena-Hard-Auto (Apache 2.0), MT-Bench (Apache 2.0), τ³-bench / tau2-bench (MIT).