anlp-a3 / README.md
winterdewdev's picture
Add model card and Part 1 checkpoints
6524139 verified
|
Raw History Blame Contribute Delete
2.44 kB
metadata
library_name: pytorch
license: mit
tags:
  - pytorch
  - mamba
  - deltanet
  - transformer
  - graph-neural-network
  - cipher

ANLP Assignment 3 checkpoints

Checkpoints for both parts of ANLP Assignment 3 (IIIT Hyderabad, 2026). All models are implemented from scratch in PyTorch. Each file holds the model weights and the full training config, and is loaded by the assignment's code.

Part 1: cipher decryption

Character-level models that decrypt a position- and context-dependent substitution cipher. The input at step i is embed(p_{i-1}) + embed(c_i) plus a sinusoidal encoding of the position in the line, and decoding is greedy. Each checkpoint is the best validation-loss checkpoint (step 50,000) of a 50k-step run with a mixed length curriculum.

File Model Parameters Char acc Word acc Norm. Levenshtein Exact lines
hybrid/best.pt (default) Interleaved MMDMMDMMDMMD Mamba / DeltaNet 6.98M 9.60% 3.44% 0.3783 535 / 1318
transformer/best.pt Decoder-only transformer, 8 layers, d 256 8.44M 9.87% 3.19% 0.3757 529 / 1318
mamba/best.pt Mamba, 16 layers, d 256, d_state 16 7.06M 9.71% 3.52% 0.3755 524 / 1318
deltanet/best.pt DeltaNet, 8 layers, 4 heads, β ∈ (0, 1) 6.90M 9.76% 3.43% 0.3755 525 / 1318

Test-set results (1,318 lines).

bash scripts/part1/test_assn.sh --heldout
bash scripts/part1/test_assn.sh --heldout --model transformer

Part 2: fake news detection on UPFD GossipCop

Graph classifiers on news propagation trees. Files are part2/{model}_{feature}_s{seed}.pt for seeds 0–4, each the best validation-F1 epoch of a 100-epoch run.

Model Features Parameters Test F1 (mean ± std, 5 seeds)
Root-only MLP bert 115,202 0.7286 ± 0.0031
GCN bert 246,530 0.9693 ± 0.0022
GraphSAGE bert 361,218 0.9637 ± 0.0023
GAT bert 247,042 0.9670 ± 0.0024
Bi-GCN bert 492,418 0.9570 ± 0.0012
GraphSAGE profile 70,146 0.9628 ± 0.0014
GraphSAGE spacy 181,506 0.9621 ± 0.0019
GraphSAGE content 185,346 0.9737 ± 0.0026
bash scripts/part2/test_gnn.sh --ta

Training logs

WandB: Part 1, Part 2