Datasets:

vinhable
/

albedo-c2-messages

messages listlengths 13 103	completion_start int64 4 80	trajectory_id stringlengths 64 64	instance_id null	task_group stringlengths 41 55	origin stringclasses 3 values	model_uri stringclasses 161 values	source stringclasses 4 values	sample_id stringlengths 37 52	sample_phase stringclasses 3 values	observation_format stringclasses 3 values	cut_point int64 1 39	split stringclasses 1 value	sampling_weight float64 0.64 5	dataset_version stringclasses 1 value	eval_run_id stringclasses 202 values	protocol stringclasses 2 values	snapshot_hash null
[ { "role": "system", "content": "You are a helpful assistant that can interact multiple times with a computer shell to solve programming tasks.\nYour response must contain exactly ONE bash code block with ONE command (or commands connected with && or \|\|).\n\nInclude a THOUGHT section before your command wher...	6	0003699be57d223282cd114eda5ac2d77af54620ec3db894b44ea1b06f337278	null	coord:mini-coder/data/train-00024-of-00060.parquet:4744	teacher	z-ai/glm-5.2	mini-coder	mini-coder/data/train-00024-of-00060.parquet:4744:2	cold	returncode	2	train	1.771875	sft-c2-teacher-first-phase-aware-v2	32652a57-a3be-4389-b8e3-f4380d971719	legacy	null
[ { "role": "system", "content": "You are a helpful assistant that can interact multiple times with a computer shell to solve programming tasks.\nYour response must contain exactly ONE bash code block with ONE command (or commands connected with && or \|\|).\n\nInclude a THOUGHT section before your command wher...	18	000574d110d39874b3dea11dd4b3364395c6f9c5f88732b463d8977a7bbcb41a	null	coord:mini-coder/data/train-00009-of-00060.parquet:429	teacher	z-ai/glm-5.2	mini-coder	mini-coder/data/train-00009-of-00060.parquet:429:8	at_edit	returncode	8	train	2.241422	sft-c2-teacher-first-phase-aware-v2	f5f477f2-f755-449a-bd2d-97886b4c339d	legacy	null
[ { "role": "system", "content": "You are a helpful assistant that can interact multiple times with a computer shell to solve programming tasks.\nYour response must contain exactly ONE bash code block with ONE command (or commands connected with && or \|\|).\n\nInclude a THOUGHT section before your command wher...	6	0010be03971dfc136e8f3a1accd4f977853a7e28f3cce0b21fd230d35a61bb4f	null	coord:mini-coder/data/train-00045-of-00060.parquet:1597	teacher	z-ai/glm-5.2	mini-coder	mini-coder/data/train-00045-of-00060.parquet:1597:2	cold	returncode	2	train	1.55925	sft-c2-teacher-first-phase-aware-v2	3387cebb-940f-4c83-934e-ef943c8d94fd	legacy	null
[ { "role": "system", "content": "You are a helpful assistant that can interact multiple times with a computer shell to solve programming tasks.\nYour response must contain exactly ONE bash code block with ONE command (or commands connected with && or \|\|).\n\nInclude a THOUGHT section before your command wher...	10	00116defb60898b43b5f1c7883ba67c9ea65c26d57312b96d7650886bfdc2d07	null	coord:mini-coder/data/train-00004-of-00060.parquet:2223	king	s3://albedo-private-models-enam/models/registrations/ac88399ae7258da203f90fc54d44ebbd6e4e35f35cf2da09de8b42ad5a3e64d1@sha256:bca411aed06e230ec16b346c4c761a44638887778543179f6176fa77a524116b	mini-coder	mini-coder/data/train-00038-of-00060.parquet:2611:4	at_edit	returncode	4	train	1.2075	sft-c2-teacher-first-phase-aware-v2	fcca6e6b-989d-4375-ba4a-713027091364	multi_reference	null
[{"role":"system","content":"You are OpenHands agent, a helpful AI assistant that can interact with (...TRUNCATED)	4	001212f0517e862d68dd69fb81e238b0518948f378ade2341cfced2dc40b01f2	null	coord:open-swe-traces/data/train-00006.parquet:1012	king	comguys/albedo-qwen3.6-35b-m33evd7@107e83dc5c164ce24f4bda6eb7c1938a281030e3	open-swe-traces	open-swe-traces/data/train-00006.parquet:1012:1	cold	openhands	1	train	0.924	sft-c2-teacher-first-phase-aware-v2	3a1882e5-ae28-46da-8178-6114bb5ae3a7	legacy	null
[{"role":"system","content":"You are a helpful assistant that can interact multiple times with a com(...TRUNCATED)	18	0013209231e7b46750b847820d02b6266eb4c03e8c5fbe605a8e3c7384e1f082	null	coord:mini-coder/data/train-00019-of-00060.parquet:996	teacher	z-ai/glm-5.2	mini-coder	mini-coder/data/train-00019-of-00060.parquet:996:8	at_edit	returncode	8	train	2.54707	sft-c2-teacher-first-phase-aware-v2	cbe5dc78-aafd-49df-911c-2351b514c7f0	legacy	null
[{"role":"system","content":"You are a helpful assistant that can interact multiple times with a com(...TRUNCATED)	14	00167b563f66c8167265798b0b8cd4ecd92918ec8911e5f303fac2ea84673ca0	null	coord:mini-coder/data/train-00047-of-00060.parquet:1355	teacher	z-ai/glm-5.2	mini-coder	mini-coder/data/train-00047-of-00060.parquet:1355:6	at_edit	returncode	6	train	1.630125	sft-c2-teacher-first-phase-aware-v2	a2b25a76-5cbd-488d-9ea8-0a27fdaa36d4	legacy	null
[{"role":"system","content":"You are a helpful assistant that can interact multiple times with a com(...TRUNCATED)	6	0016d12d70a5f88f83594e964dea8c59e4a75c50dd0fcb861fe23597c010d999	null	coord:mini-coder/data/train-00010-of-00060.parquet:2423	king	comguys/albedo-qwen3.6-35b-m33evd7@107e83dc5c164ce24f4bda6eb7c1938a281030e3	mini-coder	mini-coder/data/train-00010-of-00060.parquet:2423:2	cold	returncode	2	train	0.84	sft-c2-teacher-first-phase-aware-v2	4fcdf8d9-ef5d-454b-93f3-4d06b06cb503	legacy	null
[{"role":"system","content":"You are OpenHands agent, a helpful AI assistant that can interact with (...TRUNCATED)	12	001a27c8e7998055c34fa942f80336616471932e90121566b1ebdf17d1d91aa2	null	coord:open-swe-traces/data/train-00000.parquet:1912	teacher	z-ai/glm-5.2	open-swe-traces	open-swe-traces/data/train-00000.parquet:1912:5	pre_edit	openhands	5	train	4.385391	sft-c2-teacher-first-phase-aware-v2	f12d30b0-1227-41ed-a623-f7ef8817c2c3	legacy	null
[{"role":"system","content":"You are a helpful assistant that can interact multiple times with a com(...TRUNCATED)	4	001f45defa13701c4ef2815b356f488b0188f6ba23455c5b40f805574fd114ab	null	coord:mini-coder/data/train-00048-of-00060.parquet:1277	teacher	z-ai/glm-5.2	mini-coder	mini-coder/data/train-00048-of-00060.parquet:1277:1	cold	returncode	1	train	1.4175	sft-c2-teacher-first-phase-aware-v2	ab6a8988-0be4-4113-b26f-4ba54189c7a7	legacy	null

End of preview. Expand in Data Studio

Albedo C2 Messages

This dataset contains 16,895 training trajectories and 3,049 development trajectories, converted from the C2 teacher-first phase-aware v2 dataset. Each row has a messages list of {role, content, loss} objects, plus task, model, provenance, and sampling metadata.

The source is public Albedo evaluation artifacts with teacher, king, and challenger trajectories. The data was selected by the existing C2 quality policy; this export only changes the schema. The role/content schema was compared with vuhaian/32k_maxlift at revision 5069fbc622fa99f0bdc6ccffe30d76fa6d560b61. No examples from that dataset were added.

Load

from datasets import load_dataset

ds = load_dataset("vinhable/albedo-c2-messages", token=True)
train = ds["train"]
dev = ds["validation"]

Supervision

messages[].loss=true: assistant continuation after the original cut point.
messages[].loss=false: prefix, system, user, and environment/tool observations.
completion_start: message index where the continuation begins.
cut_point: original evaluator coordinate, not necessarily a message index.

The trainer must translate message loss flags into token labels, masking excluded tokens and padding with -100. Ordinary assistant-only loss is insufficient because the prefix also contains assistant messages. Last-response-only loss would discard intermediate continuation targets. Role names and observation strings are preserved as recorded in the original artifacts.

Sampling

train-plan.message-loss.jsonl is metadata, not a dataset split. It preserves the original C2 weighted epoch plan: 16,895 slots referencing 14,749 unique training trajectories. Use its one-based line fields to select training rows in epoch_slot order. Apply the plan once; do not also apply sampling weights to those sampled rows. Uniformly training all rows is a different recipe. The development split is task-disjoint from training.

Verification and scope

Every exported row was compared with its source for identical roles, content, supervision, and ordering. Task separation and sampling-plan identities were checked. summary.json records source and output SHA-256 hashes, row counts, and validation results.

This export does not tokenize or truncate. A 32K bound must be checked using the actual training tokenizer and chat template. End-to-end compatibility with a specific trainer still requires checking its collator and masking implementation.

Provenance and usage

Rows retain identifiers such as trajectory_id, eval_run_id, sample_id, model_uri, origin, and available source metadata. Upstream data and model terms continue to apply; this export does not grant a new blanket license. The repository is initially private.

Downloads last month: 41