kimi-v3-record

Complete run record for AgentPTB cell kimi-v3 — Kimi Code / Kimi K3 (run dir kimi-cli-fixed-v3).

A from-hour-0 redo after kimi-cli-fixed-v2 was lost to a pod eviction at h5:53. This one ran the full 100 hours (boot 2026-09-09T15:17:41Z, deadline 2026-09-13T19:17:41Z) and shut down cleanly at its deadline.

field	value
plot cell	kimi-v3
driver	Kimi Code / Kimi K3
reasoning effort	default
run boot (UTC)	2026-09-09T15:17:41Z
base model	Qwen/Qwen3.5-9B-Base (snapshot sha 68c46c4b)
training checkpoints written	92 (sft-v1 2, sft-v2 2, sft-v3 3, rl-stage1 30, rl-stage2 25, rl-stage3 15, rl-stage4 15)
submitted checkpoint	agentic-ptb/kimi-v3.h036.sft-v2.step_150
driver requests	1,073 (Kimi primary; 4 failures, zero fallback spend)

The submitted checkpoint was written at hour 36

Of 92 checkpoints spanning 83 hours, the arm submitted one written at h36.3, choosing it over a further 47 hours of its own training. It reached that conclusion deliberately, recording a plateau across seven independent attempts:

leg	swebv (stock harness, temp 1.0)
sft-v1	20%
rl-stage1 (steps 100/150/300)	16 / 19 / 17%
sft-v2 step_150	24 / 23 / 24% (n=100 ×3)
sft-v3 step_160	23%
rl-stage2	19–22%
rl-stage3	22–25%
rl-stage4 (lr 2e-6, group 16)	22%

Four GRPO stages and roughly 700 steps produced no measurable gain over the SFT checkpoint.

The cell's main finding is an eval-time one: greedy decoding

In its final ~14 hours the arm found that temperature 0.0 was worth about six points on SWE-bench-Verified on unchanged weights, because sampling at temp 1.0 caused the agent to declare success early. With a persistence prompt, greedy episodes ran 34–128 turns instead of 14.

config (same weights, sft-v2 step_150)	swebv	terminal-bench-2
stock harness, temp 1.0	24% (n=100)	1/84
stock prompt, temp 0.0	26% (n=100)	4/86
own harness + temp 0.0 (submitted spec)	30% (n=100), 34% (n=50)	5/85, 2/86, 3/63
own harness + context_window=250000	no gain — dropped	3/63

Submitted harness spec: pi + persistence/verification extra_system_prompt + allow_empty_tool_reply + temperature 0.0.

Repeat-read noise is large: identical checkpoint+config reads ranged 2–5 tb2 tasks and ±3pp swebv, so every decision above used multiple clean reads.

Independent controlled re-evaluation

The submitted checkpoint was also re-measured under the cross-cell sweep's stock pi harness (skills=[], temperature 0.6, pi 0.84.1), which measures weights rather than scaffold:

suite	n	valid	solved	pass@1	95% CI
swe-bench-verified	500	492	135	27.44%	[23.68, 31.55]
terminal-bench-2	89	87	6	6.90%	[3.20, 14.24]

Both panels clean (zero ProviderError; 1.6% / 2.2% infra loss).

Layout

path	contents
driver-session/	driver (Kimi Code) session event streams, events-*.jsonl
state/	supervisor + provider event logs, usage monitor, infra notes
evals/	eval evidence: traces and logs per panel
runs/	per-training-run configs and logs (orchestrator / trainer / inference / env)
datagen/	training data the arm generated (SFT + rejection-sampled RL data)
harness/	the arm's own harness patches, scripts and synthetic terminal tasks
NOTES.md	the arm's running research log
SUBMISSION.md	the arm's final submission write-up
goal.md	the task specification given to the driver

Weight checkpoints are published separately as agentic-ptb/kimi-v3.h*.