Clara_v4_stage3 / README.md
vphatak's picture
re-shard to 65/64/49/49; tokens and bytes identical
41f3e94 verified
|
Raw
History Blame Contribute Delete
2.62 kB
metadata
license: mit
tags:
  - clara
  - reasoning
  - pretraining
  - synthetic

Clara_v4_stage3 — targeted reasoning

172,282,390 tokens · 4 pools · pre-tokenised uint16 bins. Each pool is aimed at a measured weakness, not a guessed one. Measured on clara_1920_lockedmix (d24, n_embd 1920, dense) at step 72,000, CORE corrected:

pool tokens target task measured centered score
letter_binding_v1 75,499,631 commonsense_qa −0.0625 (below chance)
deduction_chains_v1 63,400,968 agi_eval_lsat_ar −0.0029
dyck_ops_v1 19,892,808 bigbench_dyck_languages 0.1520
repeat_ops_v1 13,488,983 bigbench_repeat_copy_logic 0.0000 (0/32)

Three of the model's four weakest CORE tasks; the fourth (bigbench_language_identification) is deliberately unserved.

Format

uint16 little-endian. BOS = 65463, EOS = 65464. Tokenizer tok_64k_singledigit, sha256 0c9c6d8c528d5aae542e27276ef9c48000dc32a8e3edab76071438d3be40d706, Digits pre-tokenizer stage present. Verify with CLARA_TOKENIZER/verify_tokenizer.py from Clara_v4_stage1 before mounting.

Dosing — for COVERAGE, not presence

Under random-offset sampling a pool served at N epochs shows only 1 − e^(−N) of itself: 1 ep = 63%, 3 ep = 95%, 5 ep = 99.3%. A targeted pool must be dosed for coverage — ≥3 epochs — or a third of it is never drawn. On a 115.8B run that is 0.196% / 0.164% / 0.052% / 0.035%.

⚠️ Ruler discipline

Every pool here targets a named benchmark task and is therefore a contamination magnet. letter_binding_v1 and deduction_chains_v1 carry import-time disjointness asserts against their probe banks. ⛔ repeat_ops_v1 shares vocabulary with 5 of repeat_copy_logic's 32 items — that is 0.156 of the task, so a gain of ≤5/32 on that task is not evidence the pool worked.

⛔ The eval bundle and every probe file are never training inputs. Pool/ruler crossing has happened four times on this program; the only safe direction is to move the pool, never the ruler.

Provenance

Entirely synthetic, generated by this lab. Items are produced by actually solving, never by templating a line per input — deduction_chains_v1 solves by constraint propagation and then backward-slices to the inferences the answer depends on.

Shards: 65 / 64 / 49 / 49. Sharded so n_shards >= world_size x 4; with shards[rank::world] a pool with fewer shards than ranks leaves some ranks with zero shards for that lane and it is silently served at a fraction of its configured dose.