{
  "paper_id": "olmoe",
  "paper_title": "OLMoE: Open Mixture-of-Experts Language Models",
  "D1": [
    {
      "id": "olmoe-D1-001",
      "claim": "OLMOE-1B-7B core Transformer architecture: 16 decoder layers, hidden_dim=2048, 16 attention heads, vocab_size=50,304 (103M vocabulary parameters), max sequence_length=4096, RoPE positional encoding with theta=10,000.",
      "source": "Section 2 (Architecture), Appendix B (Table 10)"
    },
    {
      "id": "olmoe-D1-002",
      "claim": "MoE structure: 1.3B active / 6.9B total parameters, 64 experts with 8 activated per token using dropless token choice routing applied every layer; FFN dim=1024 per expert with SwiGLU activation (no biases); no shared expert; yields 4,426,165,368 expert combinations per layer.",
      "source": "Section 2 (Architecture, MoE Configuration), Section 4.1.2, Appendix B (Table 10)"
    },
    {
      "id": "olmoe-D1-003",
      "claim": "Normalization and initialization: truncated normal weight init with std=0.02 (clamped to +/-0.06, i.e., 3x std; comparison: DeepSeekMoE uses std=0.006); RMSNorm with eps=1e-5, learnable gamma included in weight decay; QK-Norm (RMSNorm on query/key projections before attention).",
      "source": "Section 4.2.2 (Initialization), Section 4.2.3 (Normalization), Section 4.2.5 (QK-Norm), Appendix B (Table 10)"
    },
    {
      "id": "olmoe-D1-004",
      "claim": "MoE auxiliary losses: load balancing loss weight alpha=0.01, router z-loss weight beta=0.001; total training loss L = L_CE + 0.01*L_LB + 0.001*L_RZ, all computed on the same forward pass.",
      "source": "Section 2, Section 4.1.6, Section 4.1.7"
    },
    {
      "id": "olmoe-D1-005",
      "claim": "Pretraining optimizer: AdamW with beta1=0.9, beta2=0.95, epsilon=1e-8 (reduced from OLMo's 1e-5 for faster convergence); weight_decay=0.1 applied to all parameters including embeddings and RMSNorm gamma; gradient clipping by global norm at 1.0; gradient reduce and optimizer state dtypes both FP32.",
      "source": "Appendix B (Training Hyperparameters), Section 4.2.6"
    },
    {
      "id": "olmoe-D1-006",
      "claim": "Pretraining learning rate and batch: cosine schedule from peak LR=4.0e-4 to min LR=4.0e-5 with 2500 linear warmup steps; global batch_size=1024 samples (~4M tokens); checkpoint every 5000 steps (~20B tokens, approximately 1% of total pretraining).",
      "source": "Appendix B (Training Hyperparameters, Checkpoints)"
    },
    {
      "id": "olmoe-D1-007",
      "claim": "Pretraining data composition (OLMOE-MIX): 4,060B total tokens -- DCLM-Baseline 3,860B, StarCoder 101B, peS2o 57.2B, arXiv 21.1B, OpenWebMath 12.7B, Algebraic Stack 12.6B, Wikipedia/Wikibooks 3.69B. Trained for 5.133T tokens (1.3 epochs), reshuffled at each epoch boundary.",
      "source": "Section 2 (Pretraining Data)"
    },
    {
      "id": "olmoe-D1-008",
      "claim": "Data filtering rules: remove documents where any n-gram (span of 1 to 13 tokens) repeats >=32 times; StarCoder subset additionally filtered by GitHub stars >=2, most frequent word <=30% of document, top-2 most frequent words combined <=50%.",
      "source": "Section 2 (Data Filtering)"
    },
    {
      "id": "olmoe-D1-009",
      "claim": "Pretraining annealing: dataset reshuffled before annealing, learning rate linearly decayed from 4.0e-4 to 0 over final 100B tokens; post-annealing checkpoint used for adaptation (outperforms pre-annealing); load balancing loss disabled during ALL adaptation stages (SFT/DPO/KTO).",
      "source": "Section 2 (Pretraining Data), Section 4.3 (Adaptation), Appendix B"
    },
    {
      "id": "olmoe-D1-010",
      "claim": "SFT hyperparameters: max_seq_length=4096, global_batch=128 (per_device=2, gradient_accumulation=2), constant learning_rate=2.0e-5, 2 epochs, token-level loss aggregation, BF16 precision.",
      "source": "Section 4.3 (Adaptation - SFT), Appendix B"
    },
    {
      "id": "olmoe-D1-011",
      "claim": "DPO hyperparameters: global_batch=32 (per_device=1), learning_rate=5.0e-7, 3 epochs, beta=0.1; starting from SFT checkpoint, trained on UltraFeedback binarized and filtered (~60,800 preference pairs).",
      "source": "Section 4.3 (Adaptation - DPO), Appendix B"
    },
    {
      "id": "olmoe-D1-012",
      "claim": "KTO hyperparameters (alternative to DPO): RMSProp optimizer (not AdamW), 1.3 epochs (5,000 steps), other hyperparams identical to DPO; KTO performs similarly but DPO selected for final OLMOE-1B-7B-INSTRUCT due to higher AlpacaEval score.",
      "source": "Section 4.3 (Adaptation - KTO), Appendix B"
    },
    {
      "id": "olmoe-D1-013",
      "claim": "Hardware and training duration: pretraining on 256 H100 GPUs (NVLink + InfiniBand) for ~10 days; SFT on 32 H100 GPUs for 33 hours; DPO on 32 H100 GPUs for 14 hours; KTO on 8 H100 GPUs for 30 hours.",
      "source": "Appendix B (Hardware)"
    },
    {
      "id": "olmoe-D1-014",
      "claim": "Controlled MoE vs Dense experiment: MoE (1.3B active, 6.9B total, 64E/8A, FFN dim=1024) vs Dense (1.3B active, FFN dim=8192), both on 128 H100 GPUs for 130B tokens under identical config; MoE reaches dense performance with ~3x fewer tokens (~2x faster wall-clock).",
      "source": "Section 4.1.1 (Experiment Configs)"
    },
    {
      "id": "olmoe-D1-015",
      "claim": "Expert granularity ablation configs: (8E/1A, FFN=8192, 8 combos), (32E/4A, FFN=2048, 35,960 combos), (64E/8A, FFN=1024, 4.4B combos); diminishing returns beyond 32E/4A (~10% HellaSwag/MMLU gain vs 8E/1A); 64E/8A selected for final model.",
      "source": "Section 4.1.2 (Experiment Configs)"
    },
    {
      "id": "olmoe-D1-016",
      "claim": "Shared expert ablation: 31 routed + 1 shared expert (3+1 activated, 4,495 combinations) vs 32 routed (4 activated, 35,960 combinations); shared expert performed slightly worse (removes ~90% of routing flexibility), not used in final model.",
      "source": "Section 4.1.3 (Experiment Configs)"
    },
    {
      "id": "olmoe-D1-017",
      "claim": "Expert Choice vs Token Choice ablation: 8 experts, MoE every 2nd layer; TC: 2 activated per token (dropless); EC: capacity factor 2. TC outperformed EC on all tasks; EC runs ~20% faster (29,400 vs 24,400 tokens/sec/GPU) but can drop tokens. TC with load balancing selected.",
      "source": "Section 4.1.4 (Experiment Configs)"
    },
    {
      "id": "olmoe-D1-018",
      "claim": "Sparse upcycling ablation: base model OLMo-1B at 2T tokens, cloned FFN weights as 8 experts (2 activated, additional 610B tokens training); noise-based variant randomly replaces 50% of each MLP with N(0, 0.02). Not used for final model; scratch training catches up at ~500B tokens.",
      "source": "Section 4.1.5, Appendix F"
    },
    {
      "id": "olmoe-D1-019",
      "claim": "Routing analysis baselines: random baseline top-1 expert selection probability = 1.56% (1/64), random baseline top-8 = 12.5% (8/64), both derived mathematically from architecture; C4 validation sample size = 0.5% used for analysis measurements.",
      "source": "Section 5.1 (Analysis)"
    },
    {
      "id": "olmoe-D1-020",
      "claim": "OLMOE-1B-7B-0125 variant data mix: filtered DCLM 752B (50.2%), decontaminated FLAN 17.0B (16.7%), StackExchange Q&A 1.26B (2.47%), peS2o 58.6B (9.52%), Wikipedia 3.70B (3.57%), Dolmino Math 10.7B (17.5%); Dolmino-based annealing for final 100B tokens.",
      "source": "Appendix I (OLMOE-1B-7B-0125)"
    }
  ],
  "D2": [
    {
      "id": "olmoe-D2-001",
      "claim": "MoE_module(x) = sum_{i in Top^{-k}(r(x))} softmax(r(x))_i * E_i(x), where r(x) is a learned linear router, k=8, and E_i is the FFN of expert i. Implementation: compute router logits, apply softmax, select top-k, re-normalize top-k probs, compute each selected expert, weight by gate.",
      "source": "Section 2: Pretraining and Adaptation (Equation 1)"
    },
    {
      "id": "olmoe-D2-002",
      "claim": "L = L_CE + alpha * L_LB + beta * L_RZ, where alpha=0.01 (load balancing weight) and beta=0.001 (router z-loss weight). All three losses computed on the same forward pass.",
      "source": "Section 2: Pretraining and Adaptation (Equation 2)"
    },
    {
      "id": "olmoe-D2-003",
      "claim": "L_LB = N_E * sum_{i=1}^{N_E} f_i * P_i, where f_i is the fraction of tokens dispatched to expert i, P_i is the mean routing probability for expert i. Implementation: compute dispatch decisions d_{t,i}, compute f_i and P_i, then L_LB = N_E * sum_i f_i * P_i. Averaged across all MoE layers.",
      "source": "Section 4.1.6: Load Balancing Loss (Equation 3)"
    },
    {
      "id": "olmoe-D2-004",
      "claim": "L_RZ(x) = (1 / B) * sum_{i=1}^{B} (log sum_{j=1}^{N_E} exp(x_j^{(i)})) ^ 2, where x_j^{(i)} are raw router logits (before softmax). Reduces training throughput by ~2% but improves stability (fewer loss spikes) and downstream performance.",
      "source": "Section 4.1.7: Router Z-loss (Equation 4)"
    },
    {
      "id": "olmoe-D2-005",
      "claim": "RouterSaturation(t) = (1 / N) * sum_{i=1}^{N} |E_i^{(t)} intersect E_i^{(T)}| / k, measuring how early routing decisions stabilize during training. k=8 for training, k=1 also analyzed for top-1 routing. Random baseline: 1.6% (k=1), 12.5% (k=8).",
      "source": "Section 5.1: Router Saturation (Equation 5)"
    },
    {
      "id": "olmoe-D2-006",
      "claim": "ExpertCoactivation(E_i, E_j) = N_{E_i, E_j} / N_{E_i}, where N_{E_i,E_j} counts tokens where both experts are simultaneously activated, N_{E_i} counts tokens where E_i is activated. 100% means E_j always co-activates with E_i.",
      "source": "Section 5.2: Expert Co-activation (Equation 6)"
    },
    {
      "id": "olmoe-D2-007",
      "claim": "DomainSpecialization(E_i, D) = N_{E_i, D}^{(k)} / N_D, where N_{E_i,D}^{(k)} is the number of tokens from domain D for which E_i is among top-k experts, N_D is total tokens from domain D. Random baseline: 12.5% for OLMOE (8/64), 25% for Mixtral (2/8).",
      "source": "Section 5.3: Domain Specialization (Equation 7)"
    },
    {
      "id": "olmoe-D2-008",
      "claim": "VocabularySpecialization(E_i, x) = N_{x, E_i}^{(k)} / N_x, where x is a token ID (input or output), N_{x,E_i}^{(k)} counts occurrences of x routed to E_i, N_x is total occurrences of x. Distinguishes input-token vs output-token specialization.",
      "source": "Section 5.4: Vocabulary Specialization (Equation 8)"
    },
    {
      "id": "olmoe-D2-009",
      "claim": "Dropless Token Choice Routing (TC): For each token x, compute router logits r(x) of shape [N_E], select top-k experts, process x through selected experts. All tokens guaranteed k experts. Supports autoregressive generation. Used with load balancing loss.",
      "source": "Section 4.1.4: Expert Choice vs. Token Choice"
    },
    {
      "id": "olmoe-D2-010",
      "claim": "Sparse Upcycling: Clone dense FFN weights N_E times as initial expert weights, add freshly initialized router layer, continue pretraining. Upcycling is NOT used for OLMOE-1B-7B (trained from scratch instead). Scratch catches up at ~500B tokens (~25% of dense compute budget).",
      "source": "Section 4.1.5: Sparse Upcycling"
    },
    {
      "id": "olmoe-D2-011",
      "claim": "Shared Expert Architecture: One expert always active for all tokens plus k_r routed experts. Configuration: 31 routed + 1 shared (3+1 active, 4,495 combinations) vs 32 routed (4 active, 35,960 combinations). Shared expert performed slightly worse; not used in final model.",
      "source": "Section 4.1.3: Shared Experts"
    },
    {
      "id": "olmoe-D2-012",
      "claim": "Expert Choice Routing (EC): Each expert selects a fixed number of tokens from the sequence, ensuring perfect load balance. Can drop/duplicate tokens. ~20% faster training but not suitable for autoregressive generation. NOT used in OLMOE-1B-7B.",
      "source": "Section 4.1.4: Expert Choice vs. Token Choice"
    },
    {
      "id": "olmoe-D2-013",
      "claim": "Fine-Grained Expert Granularity (Section 2.2): N_total = N_experts × d_expert = constant (1.3M hidden dim); C(N_experts, k) = N_experts! / (k! × (N_experts - k)!) where k = top-k activated per layer. Use many small experts instead of few large ones. Configs: 8E/1A → C(8,1) = 8 combos, 32E/4A → C(32,4) = 35,960 combos, 64E/8A → C(64,8) ≈ 4.4B combos. Diminishing returns beyond 32E/4A. Final: 64 experts, 8 activated, FFN dim 1,024 with SwiGLU.",
      "source": "Section 4.1.2: Expert Granularity and Section 2"
    },
    {
      "id": "olmoe-D2-014",
      "claim": "Truncated Normal Initialization: Sample w ~ Normal(0, 0.02), clamp to [-0.06, 0.06] (3 * std). Improves training stability vs standard Normal init (difference clear at ~450B tokens). Used for all parameters.",
      "source": "Section 4.2.2: Initialization"
    },
    {
      "id": "olmoe-D2-015",
      "claim": "Pretraining Annealing: Reshuffle entire dataset, then linearly decay LR from 4.0e-4 to 0 over 100B tokens. Post-annealing checkpoint used for downstream adaptation (performs better than pre-annealing).",
      "source": "Section 2: Pretraining and Adaptation and Appendix B"
    },
    {
      "id": "olmoe-D2-016",
      "claim": "Repeated N-gram Dataset Filter: For each document, extract n-grams of size 1-13, count consecutive/nearby repetitions; if any n-gram repeats 32+ times, remove document. Additional StarCoder filters: >=2 GitHub stars, most-freq-word <=30%, top-2-freq <=50%.",
      "source": "Section 2: Pretraining and Adaptation"
    },
    {
      "id": "olmoe-D2-017",
      "claim": "SFT Adaptation (Appendix B): L_SFT(θ) = -E_{(x,y)∼D_SFT}[log π_θ(y|x)], trained on mixture of Tulu 2 SFT Mix, No Robots, CodeFeedback-Filtered-Instruction, MetaMathQA, Daring Anteater (filtered <4096 tokens). Token-level loss, BF16 precision, batch_size=128, constant LR=2e-5, 2 epochs, 4 H100 nodes × 8 GPUs each. No load balancing loss.",
      "source": "Section 2: Pretraining and Adaptation, Section 4.3: Adaptation Settings, Appendix B"
    },
    {
      "id": "olmoe-D2-018",
      "claim": "DPO Preference Tuning: UltraFeedback binarized and filtered (~60,800 samples). Batch 32, LR 5e-7, 3 epochs, beta=0.1, 4 H100 nodes x 8 GPUs, no load balancing loss. Starting from SFT checkpoint. DPO reference loss: L_DPO = -E[log sigma(beta * log(pi_theta/pi_ref) ratio term)].",
      "source": "Section 2: Pretraining and Adaptation, Section 4.3: Adaptation Settings, Appendix B"
    },
    {
      "id": "olmoe-D2-019",
      "claim": "KTO Preference Tuning: Same data as DPO (~60,800 UltraFeedback pairs). KTO loss: L_KTO(π_θ, π_ref) = E_{(x,y_desired,y_undesired)}[λ_d · max(0, 1 - log σ(β·log(π_θ/π_ref)(y_desired|x) - δ)) + λ_u · max(0, 1 - log σ(-β·log(π_θ/π_ref)(y_undesired|x) + δ))], where σ is the sigmoid, δ is a reference offset, β controls divergence from π_ref. RMSProp optimizer (not AdamW), 1.3 epochs (5,000 steps), other hyperparams same as DPO. Performs similarly to DPO; DPO selected for final model (higher AlpacaEval score).",
      "source": "Section 4.3: Adaptation Settings, Appendix B, Appendix F"
    },
    {
      "id": "olmoe-D2-020",
      "claim": "Pretraining Loop: Decoder-only, 16 layers, 64 experts every layer, d_model=2048, FFN dim=1024, 16 attn heads, vocab 50,304, SwiGLU, RoPE(theta=10000), full attention, RMSNorm(eps=1e-5), QK-Norm. AdamW(beta1=0.9,beta2=0.95,eps=1e-8), weight_decay=0.1(all params), peak LR=4e-4, min LR=4e-5, cosine decay, warmup 2500 steps, grad clip 1.0, seq len 4096, batch ~4M tokens, BF16 mixed precision, ZeRO FSDP.",
      "source": "Section 2, Appendix B (Table 10)"
    },
    {
      "id": "olmoe-D2-021",
      "claim": "RMSNorm: x / RMS(x) * gamma, where RMS(x) = sqrt(mean(x^2) + eps), epsilon=1e-5. Learnable gamma included in weight decay. Replaces non-parametric LayerNorm; reduces throughput by ~15% but eliminates gradient spikes.",
      "source": "Section 4.2.3: RMSNorm"
    },
    {
      "id": "olmoe-D2-022",
      "claim": "QK-Norm: Apply RMSNorm to query and key projections before attention: Q̂ = RMSNorm(Q) = RMSNorm(x W_Q), K̂ = RMSNorm(K) = RMSNorm(x W_K), then Attention(Q̂, K̂, V) = softmax(Q̂ K̂^T / √d_k) V. Uses non-parametric layer normalization (no learnable parameters) applied along the head dimension; prevents numerically large QK logits causing BF16 overflow in softmax. Improves stability and slightly improves performance; reduces throughput by ~10%.",
      "source": "Section 4.2.5: QK-Norm"
    },
    {
      "id": "olmoe-D2-023",
      "claim": "Cosine LR Schedule: LR(step) = min_LR + 0.5 * (peak_LR - min_LR) * (1 + cos(pi * step / total_steps)), with peak=4e-4, min=4e-5, 2500 linear warmup steps. After cosine, linear decay to 0 during annealing.",
      "source": "Appendix B (Table 10)"
    },
    {
      "id": "olmoe-D2-024",
      "claim": "Weight Decay: rate=0.1 applied to ALL parameters including embeddings and RMSNorm (gamma). AdamW update: theta_t = theta_{t-1} - lr * (m_hat / (sqrt(v_hat) + eps) + 0.1 * theta_{t-1}). Including RMSNorm params slightly better; including embeddings minorly better.",
      "source": "Section 4.2.3, Section 4.2.4"
    },
    {
      "id": "olmoe-D2-025",
      "claim": "AdamW with Reduced Epsilon: beta1=0.9, beta2=0.95, eps=1e-8 (reduced from 1e-5). Standard AdamW update with bias correction. Reducing eps to 1e-8 significantly improves performance (larger optimizer steps) while remaining stable.",
      "source": "Section 4.2.6: AdamW Epsilon, Appendix B"
    },
    {
      "id": "olmoe-D2-026",
      "claim": "Dataset Shuffling: At start of each epoch, randomly shuffle all pretraining samples. OLMOE trains for 1.3 epochs: shuffle at start, re-shuffle at ~4T token epoch boundary, re-shuffle again before annealing (100B tokens with linear LR decay).",
      "source": "Section 2: Pretraining and Adaptation"
    },
    {
      "id": "olmoe-D2-027",
      "claim": "SwiGLU Activation: SwiGLU(x, W, V, b, c) = Swish(x @ W + b) * (x @ V + c), with Swish(z) = z * sigmoid(z). Used in all FFN layers. No biases in OLMOE-1B-7B.",
      "source": "Appendix B (Table 10)"
    },
    {
      "id": "olmoe-D2-028",
      "claim": "Gradient Clipping (Global Norm): Compute L2 norm of all gradients; if total_norm > 1.0, scale all gradients by 1.0 / total_norm. clip_value=1.0. Used in both pretraining and adaptation.",
      "source": "Appendix B (Table 10)"
    },
    {
      "id": "olmoe-D2-029",
      "claim": "RoPE: theta=10,000. For position m and dimension pair j: freq_j = theta^{-2j/d}, apply rotation to query and key vectors. Used in all self-attention layers.",
      "source": "Appendix B (Table 10)"
    }
  ],
  "D3": [
    {
      "id": "olmoe-D3-001",
      "claim": "Pretrain OLMOE-1B-7B (6.9B total, 1.3B active, 64 experts/8 activated, dropless token choice routing, load balancing weight 0.01, router z-loss 0.001) on OLMOE-MIX for 5.1T tokens (1.3 epochs) with final 100B annealing (linear LR decay to 0). Evaluate via OLMES standard on 6 core tasks (MMLU, HellaSwag, ARC-C, ARC-E, PIQA, Winogrande) at 5-shot. Training: 256 H100 GPUs, AdamW (eps=1e-8), BF16 mixed, ZeRO FSDP, cosine LR, truncated normal init, RMSNorm, QK-Norm, weight decay 0.1 on all params. In-loop evaluation during training on MMLU Var, HellaSwag, PIQA, ARC-C, ARC-E, BoolQ, COPA, CSQA, OBQA, SciQ, SocialIQA, Winogrande, and Paloma perplexity. Baselines include Pythia-1B, OLMo-1B, TinyLlama-1B, Llama3.2-1B, DCLM-1B, OpenMoE, StableLM, DeepSeek, JetMoE, Gemma2, Qwen1.5, Llama2-7B, OLMo-7B, Mistral-7B, DCLM-7B, Llama3.1-8B, Gemma2-9B.",
      "source": "Section 2, Section 3 (Table 4), Appendix B, Appendix C"
    },
    {
      "id": "olmoe-D3-002",
      "claim": "Adapt OLMOE-1B-7B post-annealing checkpoint via two-stage pipeline: (1) SFT on Tulu 2 SFT Mix + No Robots + CodeFeedback + MetaMathQA + Daring Anteater (all <4096 tokens), token-level loss, BF16, batch 128, LR 2e-5, 2 epochs, 4 H100 nodes x 8 GPUs, no load balancing loss; (2) DPO on UltraFeedback binarized (~60,800 pairs), batch 32, LR 5e-7, 3 epochs, beta 0.1, 4 H100 nodes x 8 GPUs, no load balancing loss. KTO alternative uses RMSProp, 5,000 steps. Evaluate on MMLU (0-shot), GSM8k (8-shot CoT), BBH (3-shot), HumanEval (0-shot Pass@10), AlpacaEval 1.0 (0-shot %win), XSTest (0-shot F1), IFEval (0-shot Loose Acc). Baselines: OLMo-1B+DPO, OLMo-7B+SFT/DPO, JetMoE+SFT, DeepSeek+Chat, Qwen1.5+Chat.",
      "source": "Section 2, Section 3 (Table 5), Section 4.3 (Table 7), Appendix B, Appendix C"
    },
    {
      "id": "olmoe-D3-003",
      "claim": "Controlled MoE vs Dense comparison: MoE (1.3B active, 6.9B total, 64E/8A, FFN dim 1024) vs Dense (1.3B active, FFN dim 8192), both on 128 H100 GPUs for 130B tokens, same config except architecture. Metrics: training/validation loss, MMLU Var, HellaSwag. Report MoE reaches dense performance with ~3x fewer tokens (~2x faster wall-clock).",
      "source": "Section 4.1.1 (Figure 4)"
    },
    {
      "id": "olmoe-D3-004",
      "claim": "MoE ablation suite: (a) Expert Granularity (8E/1A vs 32E/4A vs 64E/8A), (b) Shared Experts (32 routed/4 active vs 31 routed+1 shared/3 active), (c) Token Choice vs Expert Choice routing, (d) Sparse Upcycling from OLMo-1B vs scratch, (e) Load Balancing Loss (0.01 vs none), (f) Router Z-loss (0.001 vs none). Evaluated via train/val loss, MMLU Var, HellaSwag, throughput, gradient stability, expert activation distribution. Token budgets 130B-610B per experiment.",
      "source": "Section 4.1.1-4.1.7 (Figures 4-11), Appendix F"
    },
    {
      "id": "olmoe-D3-005",
      "claim": "General pretraining ablation suite: (a) OLMOE-MIX vs Dolma 1.7 dataset, (b) Truncated normal init vs Normal init, (c) RMSNorm (parametric) vs non-parametric LayerNorm, (d) Weight decay on embeddings and RMSNorm params vs excluding them, (e) QK-Norm vs no QK-Norm, (f) AdamW epsilon 1e-8 vs 1e-5. Evaluated via train/val loss, MMLU Var, HellaSwag, PIQA, gradient norm stability, throughput. Token budgets 130B-610B per experiment.",
      "source": "Section 4.2.1-4.2.6 (Figures 12-19), Appendix F"
    },
    {
      "id": "olmoe-D3-006",
      "claim": "Adaptation ablation suite: (a) Load balancing loss during SFT+DPO (with LBL vs without), (b) Annealing checkpoint selection (pre-annealing vs post-annealing), (c) Preference optimization algorithm (DPO vs KTO with 5k/10k steps, RMS/Adam optimizers). Evaluate on 7-task benchmark (MMLU, GSM8k, BBH, HumanEval, AlpacaEval, XSTest, IFEval). Load balancing loss also measured on SFT data, GitHub, Wikipedia, C4 validation sets.",
      "source": "Section 4.3 (Tables 6-7), Appendix F (Table 14)"
    }
  ],
  "D4": [
    {
      "id": "olmoe-D4-001",
      "claim": "Pretraining pipeline ordering: Repeated N-gram Dataset Filter -> Pretraining Data Mix (OLMOE-MIX); Truncated Normal Initialization -> Model Initialization; Main Pretraining Phase (cosine LR schedule, MoE forward pass with RMSNorm + Load Balancing Loss + Router Z-Loss) -> Pretraining Annealing (dataset reshuffled, linear LR decay to 0); Annealing checkpoint -> Adaptation stages.",
      "source": "Section 2 (Pretraining and Adaptation), Section 4.2.2, Section 4.1.6, Section 4.1.7, Appendix B"
    },
    {
      "id": "olmoe-D4-002",
      "claim": "MoE forward pass dependency: Fine-Grained Expert Granularity (expert count 64 and FFN dimension 1024 design), RMSNorm, and dMoE Token Choice Routing are architectural prerequisites that feed into the MoE Module Forward Pass (Equation 1). The MoE Module's router logits are consumed by Load Balancing Loss and Router Z-Loss, which combine into Total Training Loss (Equation 2). Token Choice Routing also feeds into Load Balancing Loss computation.",
      "source": "Section 2, Section 4.1.2, Section 4.1.4, Section 4.1.6, Section 4.1.7"
    },
    {
      "id": "olmoe-D4-003",
      "claim": "Adaptation stage ordering: Pretraining Annealing produces checkpoint -> SFT Adaptation (token-level loss, constant LR 2e-5, 2 epochs) -> DPO Preference Tuning (3 epochs, LR 5e-7, beta 0.1, AdamW) OR KTO Preference Tuning (5,000 steps, RMSProp) as alternative post-SFT preference optimization. DPO was selected for final OLMOE-1B-7B-INSTRUCT. No load balancing loss during any adaptation stage.",
      "source": "Section 4.3 (Tables 6-7), Appendix B"
    }
  ]
}