task_id stringclasses 1
value | source stringclasses 1
value | score float64 0 0 | n_messages int64 5 15 | messages_json stringlengths 13.8k 49.8k |
|---|---|---|---|---|
0 | 5 | [{"role": "system", "content": "You are a data-science agent solving a scientific data-analysis task by executing Python code in a persistent sandbox.\n\n# Workflow\n1. **Inspect** the input data files first (shape, columns, dtypes).\n2. **Plan** the analytical pipeline mentally before writing code.\n3. **Execute** cod... | ||
0 | 9 | [{"role": "system", "content": "You are a data-science agent solving a scientific data-analysis task by executing Python code in a persistent sandbox.\n\n# Workflow\n1. **Inspect** the input data files first (shape, columns, dtypes).\n2. **Plan** the analytical pipeline mentally before writing code.\n3. **Execute** cod... | ||
0 | 15 | [{"role": "system", "content": "You are a data-science agent solving a scientific data-analysis task by executing Python code in a persistent sandbox.\n\n# Workflow\n1. **Inspect** the input data files first (shape, columns, dtypes).\n2. **Plan** the analytical pipeline mentally before writing code.\n3. **Execute** cod... | ||
0 | 7 | [{"role": "system", "content": "You are a data-science agent solving a scientific data-analysis task by executing Python code in a persistent sandbox.\n\n# Workflow\n1. **Inspect** the input data files first (shape, columns, dtypes).\n2. **Plan** the analytical pipeline mentally before writing code.\n3. **Execute** cod... | ||
0 | 9 | [{"role": "system", "content": "You are a data-science agent solving a scientific data-analysis task by executing Python code in a persistent sandbox.\n\n# Workflow\n1. **Inspect** the input data files first (shape, columns, dtypes).\n2. **Plan** the analytical pipeline mentally before writing code.\n3. **Execute** cod... | ||
0 | 11 | [{"role": "system", "content": "You are a data-science agent solving a scientific data-analysis task by executing Python code in a persistent sandbox.\n\n# Workflow\n1. **Inspect** the input data files first (shape, columns, dtypes).\n2. **Plan** the analytical pipeline mentally before writing code.\n3. **Execute** cod... | ||
0 | 5 | [{"role": "system", "content": "You are a data-science agent solving a scientific data-analysis task by executing Python code in a persistent sandbox.\n\n# Workflow\n1. **Inspect** the input data files first (shape, columns, dtypes).\n2. **Plan** the analytical pipeline mentally before writing code.\n3. **Execute** cod... | ||
0 | 5 | "[{\"role\": \"system\", \"content\": \"You are a data-science agent solving a scientific data-analy(...TRUNCATED) | ||
0 | 11 | "[{\"role\": \"system\", \"content\": \"You are a data-science agent solving a scientific data-analy(...TRUNCATED) | ||
0 | 5 | "[{\"role\": \"system\", \"content\": \"You are a data-science agent solving a scientific data-analy(...TRUNCATED) |
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
SFT-v3 training corpus (clean) — 138 trajectories
Best-of-pool trajectory per synth DiscoveryBench task, from 21 runs across 4 sources (8b-fold 58 / 8b-ctxgraph 44 / 8b-react 30 / 30b-ctxgraph 6), all rendered under the ctxgraph code_graph prompt. 62/200 synth tasks were EXCLUDED by a data audit (gold hypothesis references constant/missing columns in the task's own data — see audit_synth_tasks.py); corpus covers all 138 clean tasks.
Selection (threshold-light): binary gates score>0 and zero truncation retries; rank by (score desc, >=5 assistant turns preferred, duplicate-code ratio asc). Judge for selection: gpt-5-nano (Azure). Mean assistant turns 3.8.
Columns
| col | description |
|---|---|
| task_id | synth task id |
| source | teacher pool source (8b-fold / 8b-ctxgraph / 8b-react / 30b-ctxgraph) |
| score | judge HMS score of the winning trajectory on its synth task |
| n_messages | messages in the conversation |
| messages_json | FULL multi-turn conversation (system/user/assistant), no truncation |
Used by scripts/train_sft.py (LoRA r16, 2 epochs, loss on assistant tokens only). Experiment: browsecomp-ctxgraph-30b-rl (distillation v3). Status: final.
- Downloads last month
- 36