BananaMind / dsa-model-ablation

DSA ablation model (1.53M params, 3072 context)

Part of a controlled ablation: BananaMind/dsa-model-ablation (DSA) vs BananaMind/normal-model-ablation (standard attention). Same seed, data order, tokenizer, architecture and hyper-parameters. Only the attention differs.

• Attention: DeepSeek Sparse Attention (DSA): a lightning indexer (ReLU, 2 heads x dim 16) selects the top-512 keys per query; the indexer is trained with a KL loss against the (detached) main attention distribution. The first 10% of steps use dense attention (indexer warm-up), the rest use sparse attention.
• Architecture: 5 layers, hidden 128, 4 heads, SwiGLU 336, RoPE, RMSNorm, tied embeddings, vocab 4096 (custom BPE)
• Context length: 3072
• Params: 1,530,496 (1,498,496 excluding the DSA indexers)
• Data: HuggingFaceFW/fineweb-edu (sample-10BT), 2000M tokens, 1 epoch, seed 1337
• Optimiser: AdamW lr 0.002 (cosine), batch 16x3072 tokens, bf16 autocast

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "BananaMind/dsa-model-ablation"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, trust_remote_code=True)

(No KV cache is implemented; generate recomputes the full sequence each step.)

Benchmarks (0-shot, lm-evaluation-harness; acc / acc_norm where available)

Per-model raw results: benchmarks/results.json . Comparison with the other model: benchmarks/comparison.md (other repo: BananaMind/normal-model-ablation). Training curve: train_log.json .

Note: at this size models are close to chance on these benchmarks, and the benchmark prompts are shorter than the top-k (512), where DSA selects every token and behaves identically to dense attention at inference, so differences mostly reflect how training differed.