Shohin Reasoning Targets And Gaps

Bottom Line

Shohin has a healthy 125.1M-parameter training run, not a demonstrated strong reasoner. The stable loss curve and 99-100% H100 utilization establish optimization health; they do not establish state tracking, instruction following, or error correction. The live pretrain remains protected while diagnosis, curriculum work, and SFT candidates stay pinned and isolated.

Evidence

Direct interaction

The direct transcript audit and 48-case prompt matrix show this is not primarily a decoding or formatting problem:

• Raw 168k: 4/48 native Q/A, 4/48 direct-instruction, 0/48 chain-of-thought, and 5/48 one-shot cases.
• Raw 168.75k multi-turn audit: 1/6 initial, 1/6 after explicit review, and 1/6 after receiving a verified intermediate fact. The only repeated success was a simple syllogism.
• When supplied 27 * 14 = 378, the model did not reliably execute the remaining + 9 operation. Review requests repeated wrong calculations.

train/deep_interaction_audit.py extends this to the pinned 170k checkpoint. It records initial answer, review, verified-state use, model-produced compact state, and reuse of that state across arithmetic, base conversion, state updates, list and string transformation, logic, counterexample correction, and a minimal Python contract. The transcript is held-out diagnostic evidence, never training data.

Public and held-out results

• Raw 168.75k: GSM8K majority@4 5/100, greedy GSM8K 2/100, MATH-500 2/100, HumanEval 7/164, MBPP 0/100.
• V2 SFT: GSM8K majority@4 6/100, greedy GSM8K 14/100, MATH-500 6/100, HumanEval 6/164, MBPP 0/100; held-out procedural reasoning 90/800 = 11.25%.
• V4 r3: GSM8K majority@4 5/100, greedy GSM8K 14/100, MATH-500 1/100, HumanEval 2/164, MBPP 0/100. It is rejected for broad promotion, despite a useful 209/800 (26.125%) procedural held-out result. That score is above V2's 90/800 on the same evaluator, but the two pilots start from different raw checkpoints, so it is evidence to investigate rather than a clean data-only attribution. Its matched direct matrix is still only 4/48 Q/A, 4/48 direct, 4/48 CoT, and 10/48 one-shot; structured-looking derivations remain wrong on basic arithmetic and state tracking.
• V5 primitive ablation: 272/700 = 38.86% on a disjoint seven-family primitive holdout, versus raw 0/700. It proves useful transfer on a few explicit operations but is highly uneven: 100% syllogism, 88% string insertion, 47% correction, and only 2-19% on sort, base conversion, state updates, and arithmetic. Its direct audit is 2/8 initial, 2/8 review, 1/8 scaffold, and 3/8 compact-state reuse. This opens broader testing; it is not a promotion result.
• V5 broad board rejects promotion: 10/100 GSM8K majority@4, 9/100 greedy, 3/100 MATH-500, 2/164 HumanEval, and 0/100 MBPP. Its fresh seven-case operator transcript reaches 3/7 initial/review/reuse only on arithmetic, sorting, and logic; it fails base conversion, sequential state, string, and Python. The model does not actually emit its requested state= protocol, so a later correct answer is not sufficient evidence of latent-state compaction.
• V6 contract SFT makes the same distinction sharper: raw is 20/245 and V6 is 142/245 on its disjoint contract generator, but the independent deep interview is only 4/8 initial, 1/8 review, 1/8 scaffold, and 0/8 compact reuse. It has learned the generated response contracts, not robust self-correction or state compression. Its matched raw-versus-V6 prompt matrix remains an isolated gate.

External scores require matching prompts, decoding, samples, and scorer. A custom 100-example board cannot support a claim to beat another model.

Root Causes

1. No language/instruction floor in pretraining. The active equal-directory stream is about 75% math-oriented text and 25% raw code, with no material broad educational-English source. FineMath4 replayed roughly eleven times while FineMath3 had not completed one pass at 168k.
2. Old resubmissions could repeat data prefixes. Checkpoints did not record loader progress and restarted with the same seed. The forward-only stream generation fix prevents the next handoff from reusing the same ordering, but cannot make historic nominal tokens unique.
3. SFT taught style more than algorithms. The original mix was dominated by math derivations, had only hundreds of code rows, and covered too few procedural families.
4. Code was scarce and prompt-misaligned. Q/A code supervision did not match raw continuation evaluation. The prompt-boundary defect is fixed, but scale is still inadequate.
5. Latent reasoning is untrained. There is no compaction objective, trained recurrence, or test-time-loop policy in the flagship. KV caching is not self-summary. The short n_loop=2 control was stable but 1.65x slower with no demonstrated capability gain.
6. Scale is real. At about 13.3B tokens/day, 2T tokens takes roughly 150 days on the single H100. It only helps if data are unique, covered, audited, and followed by stronger post-training.

Non-Negotiable Promotion Gates

1. Data: schema, deduplication, no held-out/eval overlap, answer or execution verification where applicable, source mix, replay report, and decoded-token quality scan.
2. Atomic transfer: new primitive data must materially improve on the disjoint 3,500-row primitive holdout, per family, without a broad regression.
3. Broader reasoning: balanced held-out procedural reasoning must improve on families absent from the candidate's training source.
4. Public board: GSM8K, MATH-500, HumanEval, and MBPP cannot materially regress; code must be evaluated with a represented prompt format.
5. Direct interaction: transcripts must show correct initial, review, verified-state, and compact-state-reuse behavior. Parser-only gains do not count.
6. Generator/verifier: report greedy pass@1, oracle@K, and verifier@K. Low oracle@K means improve generation; a large oracle/verifier gap means improve verification.

Ordered Remediation

Finish evidence before touching the flagship

• Complete V4's remaining held-out/direct/verifier work only as diagnosis; V4 is already rejected for broad promotion.
• V5 has won the disjoint primitive gate but failed the public board and the contract-valid compact-state requirement. Retain it only as an atomic-transfer diagnostic; it must not affect the next broad SFT recipe.
• Admit DCLM only after its manifest and full decoded-token scan both pass.
• After the 5B DCLM pilot passes, build and scan a 25B replacement before the long language-balanced phase. The replacement must be used instead of the pilot so the streamed prefix is not double-counted.

Correct the next natural pretraining handoff

After every source passes its manifest gate, use the future-only curriculum:

• 50% audited educational English (FineWeb-Edu plus DCLM),
• 25% math/reasoning (FineMath, OpenWebMath, OpenMath),
• 25% code.

Start only from the newest numbered checkpoint with the distinct data stream, the established BS32/ACC8 single-H100 configuration, and 524,288 tokens/update. Never mutate the active job's shard list.

The loader reserves one sequence per active domain before applying weights. The future script uses floor-aware weights verified by test_domain_mix.py; weights that merely sum to 25/25/50 would silently produce the wrong batch mixture.

Scale verified post-training by missing skill

• Use solver-generated primitives for atomic operations and their disjoint set for transfer measurement.
• Scale execution-verified code in the raw continuation format used at eval.
• Scale verified math, science, logic, and procedural traces to millions of rows with source caps so small groups are not replayed into memorization.
• Admit teacher outputs only after answer checking, decontamination, source labeling, and packing analysis.
• The official Apache-2.0 OpenR1-Math default subset is a promising verified math candidate. Its intake remains gated: retain only completed per-trace Math Verify results (with LLM-judge fallback), one trace per problem, concise output, and no evaluation-prompt overlap. Its raw R1 traces are too long for this 125M-model stage, so the acceptable path is a concise source solution only if a verified R1 generation exists, the answer appears in that solution, and table/placeholder/near-answer-only artifacts are filtered. The 10k filtered pilot passed those gates; a larger candidate remains future-SFT-only.

Treat extended reasoning as separate research

First demonstrate atomic execution and useful oracle@K headroom. Then train a small controlled recurrence or compact-state objective with an equally trained non-recurrent control. A valid compaction feature needs state extraction, validation, and continuation supervision; <think>, a KV cache, or an inference-time loop does not supply that.

External Target Reality

The target is intentionally strict. The official MobileLLM-R1 140M model card reports roughly 6.2 MATH-500 and 4.1 GSM8K, while R1.5 reports 16.0 MATH-500 and 8.3 GSM8K under its own protocols. Their recipe includes far larger quality pretraining, reasoning SFT, and later on-policy distillation. These are targets to reproduce under matched protocols, not numbers to compare against a custom 100-example board.

Sources: MobileLLM-R1 140M model card, MobileLLM-R1.5 140M model card, and the official MobileLLM-R1 repository.

Current Decision

Continue the protected 1-H100 pretrain because it is stable and remains useful for the forthcoming language-balanced curriculum. Do not call it intelligent yet and do not promote V2, V4, or V5. The next claim of progress must be supported by the pinned interaction transcript, primitive transfer, valid compact-state behavior, and protocol-correct benchmark movement together.