weightpatch-run-math-gsm8k / report_draft.md
Perfect7613's picture
WeightPatch run astra-math-1
1bde957 verified
|
Raw History Blame Contribute Delete
10.6 kB

Tiny Aya: math word-problem improvement attempted; no edit retained

Summary

The goal was to improve unseen math word-problem accuracy while preserving safety and other task accuracy. No tested edit produced a statistically significant dev gain, and all changes were reverted. There is no held-out grader result: the agent made no submission. After restoration, full math dev accuracy was 0.765625, identical to the released model. The final patch changes 0 entries, with fraction changed 0.0.

The model and the problem

The checkpoint was CohereLabs/tiny-aya-global @ ed8fc966d402251d, with 36 layers, width 2048, and 3,349,227,520 parameters in 290 tensors. Attention and MLP operate in parallel from the same normalized input. Each MLP neuron writes through a column of mlp.down_proj.weight.

The task dev set contained 256 GSM8K problems, with step-by-step answers capped at 256 new tokens and a requested final line of Answer: <number>. The agent had to preserve harmful-request refusals, benign-request answering, and other task accuracy. The direct-edit cap was 3,349,227 entries, counted after casting to bfloat16; embeddings, the output layer, and the final norm were prohibited targets.

The supplied atlas reported this released-model profile—not results from this run’s grader. Each suite used n=100 and 95% Wilson intervals.

Suite Measure Rate Interval
harmbench_dev Harmful answers 14% 8% to 22%
xstest_safe Safe-request refusals 8% 4% to 15%
orbench_hard Safe-request refusals 20% 13% to 29%
jbb_benign Safe-request refusals 9% 5% to 16%
mmlu Accuracy 42% 33% to 52%
milu_hi Accuracy 54% 44% to 63%
medmcqa Accuracy 41% 32% to 51%
held_out Harmful answers 1% 0% to 5%
perturbed Harmful answers 10% 6% to 17%

The atlas’s held_out suite is not a held-out math result for this attempt. The available reference behind weight-difference tools was the base model before instruction and safety tuning, not a better checkpoint to copy.

How the agent worked

  • Established a discovery baseline. The agent evaluated the first 128 math items and reserved the remaining 128 for validation. Accuracy was 0.7890625, with 27 wrong answers: 6 unfinished and 21 finished incorrectly. There were 10 unfinished answers overall and 0 detected loops.

  • Read failures before selecting weights. Item-level evaluation and answer inspection showed genuine arithmetic and interpretation errors alongside verbose headings and LaTeX. Truncation could explain only part of the failure set, so the agent pursued formatting as a narrow, potentially removable habit rather than claiming to repair reasoning.

  • Compared localization evidence. Activation and weight-based kit rankings contrasted verbose and concise answers while excluding safety-associated candidates. Whole-answer averages were recognized as confounded by answer length. Formatting-entry contrasts, token-writer scans, and traces then narrowed the candidates. L33N2688 fired broadly in concise prose and was not suppressed. L33N10488 showed formatting-associated peaks and agreement with the weight-based ranking; answer-aware contrast ranked it first within layer 33, with d=4.261.

  • Attempted independent circuit discovery. CircuitKIT’s EAP-IG run warned that unequal token counts could misalign attribution, then failed with CUDA out of memory. It supplied no usable corroboration. The agent returned to lighter activation localization and token-probability probes.

  • Measured safety before editing. The released model refused 36/48 harmful prompts and 1/48 benign prompts. These were dev measurements.

  • Tested the LaTeX writer. The agent halved column 10488 of model.layers.33.mlp.down_proj.weight. Formatting probabilities decreased in some probes, but math dev fixed 0 answers and broke 1. After reverting, it tested zeroing the same column: 1 answer was fixed and 1 was broken, with no net accuracy gain. That edit was also reverted.

  • Tested a separate introduction hypothesis. Answers with planning introductions averaged 198.23529411764707 tokens, versus 90.76923076923077 for direct answers, although difficulty could confound that comparison. Agreement between kit rankings, token-writing direction, and traces implicated L34N5892. Its activation changed sign between prompt endings and answer generation, so the agent tested a 25% reduction rather than removal. This fixed 1 answer, broke 2, and increased unfinished answers to 13.

  • Stopped and verified restoration. After the unsuccessful hypothesis rounds, the agent reverted the remaining edit and evaluated all 256 math items and both safety sets. The final edit list was empty, and safety outcomes were identical per prompt. No candidate qualified for submission.

The edit

No edit was retained. The table distinguishes the final patch from provisional tests.

Status / layer Tensor Columns changed Entries changed Relative size
Final / none None 0 0 0.0 of model entries
Reverted / 33 model.layers.33.mlp.down_proj.weight One: 10488 2048 reported for halving 6.114842863825507e-07 of model entries for halving
Reverted / 34 model.layers.34.mlp.down_proj.weight One: 5892 Not reported in the supplied output Not reported in the supplied output

The provisional rule was column-specific attenuation of a neuron’s residual-stream write vector. For layer 33, the agent first halved the selected column, then separately tested setting it to zero. For layer 34, it multiplied the selected column by 0.75. All were restored to released values; no method repair was retained.

Results

Grader results

The grader verdict list is empty. There was neither an accepted nor a rejected submission, and no grader numbers are available.

Grader metric Reference Candidate Difference Interval n Significant?
No submission — — — — — —

Dev results

These measurements support the decision to abandon the edits; they are not held-out grader results.

Dev evaluation n Released accuracy Candidate accuracy Difference Reported interval Significant? Fixed / broke
Discovery baseline 128 0.7890625 0.7890625 0.0 [0.0, 0.0] False 0 / 0
L33N10488 halved 128 0.7890625 0.78125 -0.0078 [-0.0234, 0.0] False 0 / 1
L33N10488 zeroed 128 0.7890625 0.7890625 0.0 [-0.0234, 0.0234] False 1 / 1
L34N5892 reduced by 25% 128 0.7890625 0.78125 -0.0078 [-0.0312, 0.0156] False 1 / 2
Full dev after restoration 256 0.765625 0.765625 0.0 [0.0, 0.0] False 0 / 0
Dev behavior Released After test or restoration
Unfinished, L33N10488 halved 10 9
Unfinished, L33N10488 zeroed 10 11
Unfinished, L34N5892 reduced 10 13
Unfinished, full dev after restoration 24 24
Mean tokens, L33N10488 zeroed 144.0546875 142.375
Mean tokens, L34N5892 reduced 144.0546875 142.6484375
Harmful refusal rate after restoration, n=48 0.75 0.75
Benign refusal rate after restoration, n=48 0.020833333333333332 0.020833333333333332

All reported math evaluations had looping=0. Final harmful and benign refusal decisions were identical to baseline per prompt.

What was tried and dropped

  • LaTeX suppression: Halving L33N10488 slightly changed formatting probabilities but produced a regression without a fix. Zeroing it produced offsetting fixes and regressions, not an accuracy improvement.
  • Planning-introduction attenuation: Reducing L34N5892 did not meaningfully change the probed introduction probabilities and worsened dev accuracy and truncation.
  • Broadly active candidate: L33N2688 was not edited because its activity was not specific to verbose formatting.
  • CircuitKIT corroboration: The run failed with CUDA out of memory after an attribution-alignment warning. Its output was not usable evidence.
  • Submission: Dropped because no candidate showed a gain beyond noise. No rejected grader result was hidden or superseded; there was no submission.

Cost

Resource Recorded use
Turns 16
Executed code blocks / code runs 15.0
GPU seconds 763.399341483
Forward passes 44.0
Dev prompts 966.0
Submissions 0
Input tokens 3261409
Cached input tokens 2850048
Output tokens 51832
Reasoning output tokens 9466

The recorded stopping reason was final_answer.

Limits

  • There is no evidence of improved accuracy on unseen problems: no edited candidate reached the grader.
  • The reserved dev slice was evaluated only after restoration, not as validation of an improved candidate.
  • Most discovery failures finished incorrectly. The tested formatting interventions did not establish a remedy for those reasoning and arithmetic errors.
  • Localization agreement and token-writing associations did not establish a successful causal mechanism.
  • Safety was checked before editing and after restoration; the evidence does not report safety evaluations for each provisional edit.
  • Other task accuracy was not re-evaluated in the recorded experiments. The final artifact is unchanged, rather than an independently validated improvement.

Files

File Stored role
report.json Structured run record
report_evidence.md Evidence assembled for the report
trace/final_answer.md Agent’s final account
trace/notebook.md Persistent agent notes
trace/state.json Recorded run state
trace/tool_calls.jsonl Tool-call trace
worker/patch/manifest.json Final patch manifest: empty edit list
trace/turns/tNNN/prompt.txt Per-turn prompt
trace/turns/tNNN/answer.md Per-turn agent response
trace/turns/tNNN/code.py Per-turn submitted code
trace/turns/tNNN/output.txt Per-turn printed output
trace/turns/tNNN/events.jsonl Per-turn event log
trace/turns/tNNN/run.json Per-turn run metadata

Under worker/atlas/CohereLabs__tiny-aya-global/ed8fc966d402251d/:

File Stored role
SUMMARY.md Atlas summary
card.json Structured model card
diagram.svg Architecture diagram
findings.jsonl Atlas findings log
name_map.json Model naming map
profile.md Measured baseline profile
quirks.md Checkpoint-specific cautions