KevinDavidHayes/long-context-baseline-bakeoff

Long-Context Data-Selection Bake-off — Shared Candidate Pool

The shared 16K candidate pool for comparing long-context data-selection methods on equal footing. Every method (AttentionSpan, LongAttn, LongProc, ProLong, perplexity, ...) scores the same 14,300 documents, picks its own top-800 under the same split, then trains Llama-2-7B + 16K LoRA and evaluates on HELMET.

Files

File	Description
candidate_pool_16k_scored.parquet	The shared pool — 14,300 docs, each ≥16,384 Llama-2 tokens, 7 domains.
selections/as_positive_800.txt	Reference: AttentionSpan global-top-800 (doc_id list).
selections/as_balanced_800.txt	Reference: AttentionSpan domain-balanced top-800.
selections/random_800_seed42.txt	Shared random-800 baseline (seed 42).
scripts/longattn_score_full_pool.py	LongAttn scorer/selector that reads this pool.
pool_manifest.json	Provenance, score definition, split spec.

Pool schema

Column	Description
doc_id	sha1(text) — stable join key across all selectors
text	Full document text (score this)
source	Domain: code / web / arxiv / encyclopedia / books / government / legal
token_length	Original (Qwen) token length
llama2_token_length	Llama-2 token length (all ≥16,384)
sequence_avg_median_lookback	Raw attention lookback (Qwen2.5-Coder-7B, layer 27)
long_token_ratio	AttentionSpan score = sequence_avg_median_lookback / 16384

Split (identical for every selector)

• Score all 14,300 docs, take the top-800 by the selector's own score.
• min_tokens = 16,384 (Llama-2 tokenizer); seed = 42 for any random draw.
• Reuse random_800_seed42.txt as the common baseline row across all methods.

Training + eval recipe (must match)

Llama-2-7B, 16K via linear RoPE factor 4; LoRA r=16 α=32 on q/k/v/o; 1 epoch, effective batch 8; HELMET 16K, 22 subtasks (Recall 8 + RAG 8 + LongQA 6), 100 samples/subtask; paired bootstrap B=5,000.