bird35/quasar-sn24-teacher-cache

Quasar SN24 Teacher Cache (Top-128 Sparse Logprobs)

Teacher distillation cache for Bittensor Subnet 24 (Quasar). Each .pt file contains Qwen/Qwen3.5-4B sparse top-128 logprob outputs over ClimbMix prompts, produced by the validator's official build_cache.py pipeline.

Contents

• —cache100/teacher_cache_blkN.pt — 59 individual block-seeded caches, ~190 prompts each, drawn from distinct ClimbMix shards (shard = seed % 6542). Format: {items: [{full_ids, prompt_len, indices[seq,128], values[seq,128]}, ...]}
• —teacher_cache_8252640.pt — the real production block 8252640 (hash 0x284034b75424cc4b...), 184 prompts from shard 5031.
• —teacher_cache_mega_fresh.pt — combined evalfmt format (10,694 prompts from 58 blocks + 8252640) ready for training. Format: {full_sequences, prompt_lens, teacher_logits, prompts_hash, n_prompts}.

Usage

Drop-in cache for off-policy distillation training (train_v31.py, train_v31_rkl.py). Each prompt comes with full token sequence (prompt + teacher generation) plus top-128 logprob distribution for KL loss computation.

Teacher: Qwen/Qwen3.5-4B with temperature=0.7, top_p=0.9, max_new_tokens=768, logprobs=128.