mse_validation/notes/morphology-experiment-2026-05-05.md

Morphology-Aware Synthetic Filler — SWE-bench MSE Validation

Date: 2026-05-05 Status: chars/token fixed; turn-10+ E2EL gap partially improved but not closed

Hypothesis

SWE-bench distributional MSE E2EL gap (turn 10-19: +45.5%) is partly caused by synthetic prompts having ~8.4 chars/token (English filler) vs REAL ~3.8 chars/token (code/tool-output text). Matching the chars/token ratio should reduce request-preprocessing and prefix-handling differences if character density is a causal driver rather than just a correlated symptom.

Method

Source-locked MSE+REAL pair on same vLLM instance, same 40 SWE sessions. Two conditions:

Condition	Host	DISTRIBUTIONAL_SYNTHETIC_STYLE	DISTRIBUTIONAL_TARGET_CHARS_PER_TOKEN
English (baseline)	h100 (gpu-13)	(default)	(default ~8.4)
Code (experiment)	h100-2 (gpu-15)	code	3.8

Both runs: SESSIONS=40, CONC=5, PORT=8091/8092, GPU_MEM=0.75, MAX_LEN=32768, SOURCE_SESSION_IDS_FILE source-locked to same 40 SWE session IDs extracted from the English REAL run.

Source IDs extracted from:

results/mse_validation_source_locked_pair/h100_swebench_c5_s40/Llama-3.1-8B_tp1_vllm/swebench-multiturn-short_conc5.json

Command:

DISTRIBUTIONAL_SYNTHETIC_STYLE=code \
DISTRIBUTIONAL_TARGET_CHARS_PER_TOKEN=3.8 \
SOURCE_SESSION_IDS_FILE=/tmp/swebench_source_locked_ids.txt \
bash scripts/run_mse_validation.sh /data/models/Llama-3.1-8B-Instruct 1 Llama-3.1-8B vllm \
  swebench 5 results/mse_validation_morphology/h100_swebench_c5_s40_codechars \
  /home/kevinlau/miniconda3/envs/vllm/bin/python 0.75 32768

Results

Criterion 1: chars/token ratio

Condition	Median chars	Median tokens	chars/token	vs REAL
English MSE	57,004	6,798	8.39	2.2×
Code MSE	25,037	6,798	3.68	1.0×
REAL	24,618	6,472	3.80	—

Per-bin prompt_chars Δ (Code MSE vs REAL): −2% to +6%. Target hit.

Criterion 2: E2EL p50 per turn bin

Turn bin	English MSE	English REAL	English Δ	Code MSE	Code REAL	Code Δ	Improvement
00–04	214ms	207ms	+3.2%	221ms	217ms	+1.9%	+1pp
05–09	348ms	340ms	+2.2%	343ms	339ms	+1.0%	+1pp
10–19	675ms	464ms	+45.5%	670ms	511ms	+31.1%	+14pp
20–29	2,267ms	1,957ms	+15.8%	2,128ms	1,893ms	+12.4%	+3pp

Input token Δ per bin: identical between conditions (+2.3% to +4.0% across bins). Session overlap: 40/40 (Jaccard 1.00).

Aggregate

Metric	English Δ	Code Δ
TTFT p50	+46.8%	+23.8%
TPOT p50	+8.9%	+14.7%
E2EL p50	+27.0%	+21.6%

Verdict

Chars/token ratio is a partial cause, not the whole cause. Matching the morphology from 2.2× to 1.0× closed ~14pp of the 45% turn-10–19 E2EL gap, but the remaining +31% gap is outside the observed noise floor. The residual likely stems from content structure differences beyond simple character density: real SWE traces contain recurrent boilerplate (traceback frames, pytest output, git diffs), while code-like filler generates independent random fragments. The experiment does not establish a different KV-cache block layout; same token count should imply similar sequence length and KV block count.

Confounding factor

The English and Code runs used different GPU hosts:

• English: h100 / gpu-13
• Code: h100-2 / gpu-15

REAL baselines differ (E2EL 515ms vs 551ms), so ~5–10pp of the residual gap may be cross-host noise. A same-host A/B test would be needed to isolate this.

Result Files

Morphology (code-like filler) run

results/mse_validation_morphology/h100_swebench_c5_s40_codechars/Llama-3.1-8B_tp1_vllm/
  swebench-multiturn-mse-short_conc5.json           ← MSE (code filler)
  swebench-multiturn-mse-short_conc5_per_turn.json
  swebench-multiturn-short_conc5.json               ← REAL
  swebench-multiturn-short_conc5_per_turn.json

English filler baseline (same-instance source-locked pair)

results/mse_validation_source_locked_pair/h100_swebench_c5_s40/Llama-3.1-8B_tp1_vllm/
  swebench-multiturn-mse-short_conc5.json           ← MSE (English filler)
  swebench-multiturn-short_conc5.json               ← REAL

Source session IDs

/tmp/swebench_source_locked_ids.txt   (40 IDs, on h100-2)

Next Probe

Recurrent content filler. Instead of independent code fragments per request, reuse actual prefix text chunks from SWE real traces. If the gap closes, recurrent boilerplate/content structure is the root cause; if not, the remaining error is somewhere else in the serving input path or runtime variance rather than simple chars/token morphology.

Prefix-Aware Synthetic Follow-Up

vLLM automatic prefix caching reuses KV blocks when later requests share an exact token prefix with previous requests. The distributional generator already builds a growing transcript within each synthetic session, so each turn is prefix-eligible relative to the previous turn in the same session. What it did not model was the shared cross-session harness prefix that real SWE/TerminalBench traces usually have.

Local change added on 2026-05-05:

• DISTRIBUTIONAL_PREFIX_AWARE=1 enables a fixed shared system-prefix for distributional synthetic sessions.
• DISTRIBUTIONAL_SHARED_PREFIX_TOKENS=1024 sets the content-token target for that shared prefix.
• DISTRIBUTIONAL_PREFIX_BLOCK_SIZE=16 aligns the shared prefix target to vLLM-style cache blocks.
• PREFIX_AWARE_SYNTHETIC=on in scripts/run_mse_validation.sh enables this only for the MSE/distributional side of a paired run.

This preserves sampled total-context token targets by subtracting the shared prefix from the first-turn synthetic user payload. The result is an APC-aware ablation: same token budget, same morphology controls, but with cross-session shared prefix structure.

Suggested next run:

PREFIX_AWARE_SYNTHETIC=on \
DISTRIBUTIONAL_SYNTHETIC_STYLE=code \
DISTRIBUTIONAL_TARGET_CHARS_PER_TOKEN=3.8 \
SHARED_PREFIX_TOKENS=1024 \
SOURCE_SESSION_IDS_FILE=/tmp/swebench_source_locked_ids.txt \
bash scripts/run_mse_validation.sh /data/models/Llama-3.1-8B-Instruct 1 Llama-3.1-8B vllm \
  swebench 5 results/mse_validation_prefix_aware/h100_swebench_c5_s40_codechars_sharedprefix \
  /home/kevinlau/miniconda3/envs/vllm/bin/python 0.75 32768

Code

• inference-benchmark/src/workloads/distributional.py:136 — DISTRIBUTIONAL_SYNTHETIC_STYLE / DISTRIBUTIONAL_TARGET_CHARS_PER_TOKEN
• inference-benchmark/src/workloads/distributional.py:147 — DISTRIBUTIONAL_PREFIX_AWARE / shared-prefix controls
• inference-benchmark/src/workloads/distributional.py:381 — _synthetic_text() style dispatch
• inference-benchmark/src/workloads/distributional.py:483 — _calibrated_morphology_text() with target_chars_per_token