oracle-lens-data / README.md
agu18dec's picture
README: cross-link data <-> AR <-> AO repos, refresh counts
c260146 verified
|
Raw
History Blame Contribute Delete
6.86 kB
metadata
license: other
license_name: research-only
license_link: LICENSE

oracle-lens-data — Inverted OLens (AR + AO)

All data for the Inverted OLens program: fully fresh on-policy rollouts from Qwen/Qwen3.6-27B, the whiteners fit on them, and the AO pool. Checkpoints live in agu18dec/oracle-lens-ar-checkpoints (AR) and agu18dec/oracle-lens-ao-checkpoints (AO).

Activations (AR training pairs) are deliberately NOT stored here — they are ~174 KB each (17 layers x 5120 x bf16) and are deterministically re-derivable from the rollouts below. See Regenerating activations.

Layout

path contents
rollouts/chat/ on-policy assistant rollouts — 1,485,997 convs / 798,403,679 output tokens (12 shards + reports/)
rollouts/pt/ on-policy pretraining-text rollouts — 991,520 convs / 680,133,035 output tokens (13 shards + reports/)
whiteners/chat/, whiteners/pt/ per-layer mean+cov (fp64-accumulated), 17 layers per cell, >=1.04M rows each
ao/pool/ AO crop pools, fingerprinted: pool_iolens (527,160 crops) + eval_pool_iolens (13,137), and the exclusion-deduped extension pool_iolens_ext1 (939,113) + eval_pool_iolens_ext1
ao/arout/<ar-run>/<rung>/ AR reconstructions the AO trains on — k=4 seeded layers/crop, layers ≥ 20, self-describing metadata (ao_layers, pick seeds)
ao/runs/ frozen injection scale (scale_iolens_chat_final.json: alpha 16000 / scale 64.559), arout gate + sweep records
meta/ splits, seed reports, exact datagen counts, gate reports (prelaunch_*.json), per-shard capture manifests

Rollout shard schema (safetensors)

tensor dtype meaning
ids int32 [total] prompt then output tokens, per conversation, concatenated
offsets int64 [n+1] conversation i = ids[offsets[i]:offsets[i+1]]
prompt_len int32 [n] first prompt_len[i] tokens are the prompt; the rest is the model's own generation
seed_hash int64 [n] blake2b-8 of the seed key (freshness ledger)
split_id int8 [n] 0 ar_train / 1 ao_train / 2 ao_val / 3 eval — assigned from seed content hash before generation
out_logprob fp16 [total_out] engine logprob of each sampled output token (ragged, conversation order)

Shard metadata (meta key) carries model id, mode, engine + version, tokenizer sha, sampling params and exact token counts.

How the data was made

  • chat: WildChat-1M user turns -> apply_chat_template(..., enable_thinking=False) -> SGLang sampling at temp 1.0, top_p 1.0, max_new 1536.
  • pt: FineWeb-Edu documents -> first 256 tokens as a raw prefix (NO chat template) -> base-style continuation, temp 1.0, top_p 1.0, max_new 1024.
  • Splits are a pure function of seed content hash, assigned before generation and conversation-cohesive (all turns of a conversation land on one side).
  • Degenerate outputs (empty/EOS-only, or an immediately repeating 20-gram) were dropped and counted per shard. Verified gates: tokenizer identity, byte-exact prompt renders (1000/1000 per cell), engine-vs-HF teacher-forced logprob parity, per-shard length/degeneracy, seed freshness. Reports in meta/.

Regenerating activations

Activations are captured by a single teacher-forced forward over each stored conversation, with spans carved only inside the generated region. Everything is seeded, so the output is bit-identical to what the program trained on.

git clone https://github.com/camilablank/global-workspace && cd global-workspace
bash bootstrap_runpod.sh --all                 # env + torch pin + serving envs
source scripts/cluster/env.sh
hf download agu18dec/oracle-lens-data --repo-type dataset \
    --include 'rollouts/chat/*' --local-dir $OLA_ROOT/rollouts_dl

# one GPU, one rollout shard -> multilayer_v1 pair shards
CUDA_VISIBLE_DEVICES=0 uv run --no-sync python scripts/ola/iolens_capture_pairs.py \
    --mode chat --rollout-shard 0 \
    --rollouts-dir rollouts_dl/rollouts/chat \
    --out-dir ml_pairs_chat --train-frac 0.14        # slice of ar_train; skip-cursor for waves

What that produces, per pair:

  • input = a span of N ~ uniform{1..32} tokens taken at a sampled position inside the assistant/continuation region. Spans within a conversation are mutually disjoint (no span is a prefix or subset of another, no activation is shared between pairs), ~35 spans/conv.
  • target = the residual stream at prev_pos = span_start - 1 — the state the model was in just before emitting that span — at layers (0, 4, 8, ..., 60, 63), stored [17, 5120] bf16.
  • Conversations whose output re-emits a <think> block (chat, ~8%) or loops are skipped; mojibake is kept (it is inherited from source documents and is faithful on-policy text).

Useful flags: --carve-mode octave (legacy long-span cascade, for >32-token crop-source experiments), --train-frac-skip (disjoint later waves), --batch-tokens/--batch-rows (capture throughput).

Fit whiteners on the pairs you just captured (needed for the AR loss and every FVE number):

CUDA_VISIBLE_DEVICES=0 uv run --no-sync python scripts/ola/iolens_fit_whitener.py \
    --pairs-dir ml_pairs_chat --out-prefix whitening_chat --n-shards-use 12

...or just use the published whiteners/{chat,pt}/ (fit on >=1.04M rows/layer; null-predictor FVE ~ 1/5120, i.e. random vectors score exactly chance in this basis).

Which data goes to which trainer

consumer source why
AR (assistant cell) rollouts/chat split ar_train -> pairs learns to invert assistant text
AR (PT cell) rollouts/pt split ar_train -> pairs learns to invert document continuations
AR eval (both) split eval -> pairs (true activations) held-out FVE / retrieval
AO (all runs) rollouts/chat split ao_train -> ao/pool the AO always reads assistant text, whichever AR produced its activations — that is what makes chat-AR vs PT-AR a controlled comparison
AO val conversation-level split of the same pool (seed 1234) no window/prefix overlap with AO train

Hygiene that is enforced, not assumed: AR and AO conversations are disjoint by pre-generation split; every AO crop is additionally checked against a hash of every AR span prefix (31,889 crops dropped, 0 collisions remain); AO windows are non-overlapping with one crop length each; audit_diversity refuses any configuration where a text would be a target more than 2.5x/epoch by default (the k=4 program overrides to 4.0 with --max-repeat, watching val CE for a memorization inflection).