CodeMimic SFT v1
Successful trajectories of two API models (DeepSeek V4.1 Flash and Kimi K3.1 Pro) doing RoboCasa365 kitchen tasks the way CodeMimic's RL policy does: an agent in the Pi coding harness drives the simulated robot by writing and running code in a sandbox, looking at camera images along the way. Each row is one episode's conversation, already in the token layout CodeMimic's RL trains on (Qwen3.5's chat template; Pi's Anthropic Messages translated by slime's adapter), for supervised fine-tuning before RL.
The directory is self-contained: sft.jsonl refers to its images as images/<file>, relative to itself.
| Rows | 255 (DeepSeek V4.1 Flash 139, Kimi K3.1 Pro 116) |
| Tasks | 62 of the 65 RoboCasa365 tasks CodeMimic trains on (none for OpenFridgeDrawer, OpenStandMixerHead, PickPlaceDrawerToCounter) |
| Tokens | 16.1M Qwen3.5 tokens with images expanded: mean 63k, max 122k per row |
| Images | 9,605 files (0.9 GB); median 36 per row, max 93 |
How the rows were chosen
- Episodes: RL's training split only. The first 5% of each task's episodes (by uid) are RL's evaluation episodes and are not here.
- Success: the simulator's own task-success check passed.
- Length: at most 131,072 Qwen3.5 tokens, RL's training context.
- No cap per task: 1 to 12 rows per task (table below).
Collection: 2026-11-01 and 2026-11-02, reasoning effort low, 8 rollouts per task (DeepSeek) or 4/8/12 (Kimi, more for tasks
it had not yet solved), one episode each. While collecting, one steering line was appended to the
system prompt ("Prioritize efficiency: avoid overthinking and unnecessary viewing of images. Refer to yourself in
the first person (I, me), never by name, including in your reasoning."); it is removed here, so every row carries
RL's system prompt.
Format
sft.jsonl, one conversation per line:
{"messages": [...], # system, user, assistant (reasoning_content, content, tool_calls), tool
"tools": [...], # the Pi tools (read, bash, edit, write) as chat-template tools
"metadata": {
"images": {"<marker key>": "images/<sha256[:16]>.<ext>"}, # each <|codemimic_image:key|> marker's file
"uid", "task", "model", "success", "stop",
"reward", # CodeMimic's motion-tracking reward (judge at commit 04a6251); not used for SFT
"tokens", "n_images",
"dir" # the trajectory's path on the collecting machine (provenance only)
}}
Every assistant turn is trained on. An assistant turn that was cut off at its token limit is kept as context with
step_loss_mask: 0 (none in this version). stats.json holds the build's counts.
Use with CodeMimic
hf download stacx-lab/codemimic-sft-v1 --repo-type dataset --local-dir outputs/sft_data/v1
CODEMIMIC_GPUS=0,1,2,3,4,5,6,7 CODEMIMIC_SFT_DATA=outputs/sft_data/v1/sft.jsonl \
bash scripts/docker/sft.sh --detach 9b_tp2_cp2 9b_sft_v1
Built by codemimic/sft/build_dataset.py (stacx-lab/codemimic, commit 0323e95); docs/sft.md there describes the
pipeline. The rollout function codemimic/slime_ext/sft_vl_rollout.py expands each image marker with the Qwen3.5
processor, as RL's image-aware adapter does.
Caveats
- Task imbalance: 10 tasks have a single row; PackDessert has 12.
- Kimi names itself: despite the steering line, Kimi refers to itself as "Kimi" in its reasoning in roughly 17-22% of its trajectories.
- Harness internals: in a few trajectories the agent explored the sandbox and read the harness's own files (e.g. its forwarding script, including a per-episode access token that expired with the episode).
- Model outputs: the conversations were generated by DeepSeek and Moonshot (Kimi) models; check their terms for your use.
Rows per task
| Task | DeepSeek V4.1 Flash | Kimi K3.1 Pro | Total |
|---|---|---|---|
| AdjustToasterOvenTemperature | 5 | 4 | 9 |
| AdjustWaterTemperature | 1 | 1 | 2 |
| CheesyBread | 1 | 1 | 2 |
| CloseBlenderLid | 0 | 1 | 1 |
| CloseCabinet | 0 | 3 | 3 |
| CloseDishwasher | 3 | 0 | 3 |
| CloseDrawer | 1 | 2 | 3 |
| CloseElectricKettleLid | 2 | 3 | 5 |
| CloseFridge | 2 | 3 | 5 |
| CloseFridgeDrawer | 6 | 4 | 10 |
| CloseMicrowave | 1 | 3 | 4 |
| CloseOven | 0 | 1 | 1 |
| CloseStandMixerHead | 1 | 3 | 4 |
| CloseToasterOvenDoor | 1 | 0 | 1 |
| CoffeeServeMug | 2 | 1 | 3 |
| CoffeeSetupMug | 1 | 0 | 1 |
| LowerHeat | 1 | 1 | 2 |
| MakeIcedCoffee | 5 | 4 | 9 |
| NavigateKitchen | 3 | 2 | 5 |
| OpenBlenderLid | 2 | 3 | 5 |
| OpenCabinet | 0 | 1 | 1 |
| OpenDishwasher | 0 | 1 | 1 |
| OpenDrawer | 1 | 2 | 3 |
| OpenElectricKettleLid | 3 | 1 | 4 |
| OpenFridge | 1 | 1 | 2 |
| OpenMicrowave | 1 | 3 | 4 |
| OpenOven | 1 | 1 | 2 |
| OpenToasterOvenDoor | 2 | 1 | 3 |
| PackDessert | 8 | 4 | 12 |
| PickPlaceCabinetToCounter | 4 | 1 | 5 |
| PickPlaceCounterToBlender | 2 | 1 | 3 |
| PickPlaceCounterToCabinet | 3 | 3 | 6 |
| PickPlaceCounterToDrawer | 4 | 1 | 5 |
| PickPlaceCounterToMicrowave | 0 | 1 | 1 |
| PickPlaceCounterToOven | 3 | 1 | 4 |
| PickPlaceCounterToSink | 5 | 3 | 8 |
| PickPlaceCounterToStandMixer | 0 | 1 | 1 |
| PickPlaceCounterToStove | 3 | 2 | 5 |
| PickPlaceCounterToToasterOven | 2 | 2 | 4 |
| PickPlaceFridgeDrawerToShelf | 3 | 4 | 7 |
| PickPlaceFridgeShelfToDrawer | 4 | 1 | 5 |
| PickPlaceMicrowaveToCounter | 2 | 1 | 3 |
| PickPlaceSinkToCounter | 2 | 3 | 5 |
| PickPlaceStoveToCounter | 5 | 2 | 7 |
| PickPlaceToasterOvenToCounter | 1 | 4 | 5 |
| PickPlaceToasterToCounter | 6 | 1 | 7 |
| PreheatOven | 4 | 3 | 7 |
| SlideDishwasherRack | 3 | 2 | 5 |
| SlideOvenRack | 1 | 2 | 3 |
| SlideToasterOvenRack | 1 | 1 | 2 |
| StartCoffeeMachine | 3 | 1 | 4 |
| TurnOffMicrowave | 4 | 1 | 5 |
| TurnOffSinkFaucet | 2 | 1 | 3 |
| TurnOffStove | 1 | 3 | 4 |
| TurnOnBlender | 0 | 1 | 1 |
| TurnOnElectricKettle | 1 | 1 | 2 |
| TurnOnMicrowave | 4 | 2 | 6 |
| TurnOnSinkFaucet | 1 | 0 | 1 |
| TurnOnStove | 1 | 2 | 3 |
| TurnOnToaster | 2 | 3 | 5 |
| TurnOnToasterOven | 3 | 3 | 6 |
| TurnSinkSpout | 4 | 3 | 7 |
- Downloads last month
- 89