ANLP Assignment 2 - Monsoon 2026

Author: Krithik Kambhampati (krithik3001)
Institution: IIIT Hyderabad

Checkpoints, tokenizer, plots and evaluation results for Assignment 2. All numbers below are generated from the result JSON files in this repository.

WandB Projects

Part 1: Mixture of Experts (test split of belumind/en-vi-ja-curated-500k-triplets)

Variant Total Params Active Params Test Loss Test PPL Test BLEU Checkpoint
dense 27,422,720 27,422,720 2.3242 10.22 24.57 part1/best_dense.pt
moe_top1 27,444,224 17,988,608 2.3633 10.63 24.35 part1/best_moe_top1.pt
moe_top2 27,444,224 21,140,480 2.3427 10.41 24.50 part1/best_moe_top2.pt
moe_shared 27,441,152 21,137,408 2.2974 9.95 25.01 part1/best_moe_shared.pt
moe_active_match 40,039,424 27,438,080 2.3122 10.10 24.42 part1/best_moe_active_match.pt

Routing heatmaps: assets/routing_heatmaps/.

Part 2: Custom Optimizers (browndw/human-ai-parallel-corpus, final 1.0x milestone)

Optimizer LR Tokens Val Loss Val PPL Continuation BLEU Checkpoint
adamw 0.0003 52,152,638 3.5396 34.45 1.12 part2/adamw_1.0x.pt
nadamw 0.0003 52,152,638 3.5231 33.89 1.15 part2/nadamw_1.0x.pt
lion 0.0003 52,152,638 3.7967 44.56 0.55 part2/lion_1.0x.pt
muon 0.02 52,152,638 3.3953 29.82 1.20 part2/muon_1.0x.pt
sophia 0.0003 52,152,638 4.4071 82.03 0.47 part2/sophia_1.0x.pt

Training dynamics: assets/part2_dynamics/.

Part 3: Decoding Strategies (EleutherAI/Pythia-160M on hamishivi/ROCStories)

Reference (human continuation) perplexity: 25.03. Greedy == beam(W=1) on 100.0% of samples.

Strategy PPL Token Acc (%) F1 (%) BLEU Avg Gen Tokens ms / sample ms / token
greedy 2.18 1.43 13.27 0.44 64.0 274.1 4.28
top_k_10 7.70 1.48 19.60 0.50 63.6 264.8 4.16
top_k_50 13.77 1.32 18.74 0.34 63.2 246.6 3.90
top_p_0.7 10.30 1.38 18.87 0.32 63.3 243.9 3.85
top_p_0.9 24.45 1.17 17.42 0.30 62.9 243.2 3.87
top_p_0.95 34.56 1.11 16.74 0.20 63.2 244.6 3.87
beam_width_1 2.18 1.43 13.27 0.44 64.0 245.9 3.84
beam_width_2 1.94 1.22 11.50 0.30 63.9 250.9 3.93
beam_width_4 1.80 1.17 10.79 0.29 63.6 256.8 4.04

Raw generations: assets/part3/rocstories_generations.json.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support