sergiopaniego/Qwen3.5-2B-SmolDataEnvs-SFT-500
Image-Text-to-Text • 2B • Updated • 7
Datasets:
messages listlengths 7 37 | tools listlengths 10 10 | task_name stringlengths 20 22 | n_turns int64 4 19 |
|---|---|---|---|
[
{
"role": "system",
"content": "You are opencode, an interactive CLI tool that helps users with software engineering tasks. Use the instructions below and the tools available to you to assist the user.\n\nIMPORTANT: You must NEVER generate or guess URLs for the user unless you are confident that the URLs ar... | [
{
"type": "function",
"function": {
"name": "bash",
"description": "Executes a given bash command in a persistent shell session with optional timeout, ensuring proper handling and security measures.\n\nBe aware: OS: linux, Shell: bash\n\nAll commands run in the current working directory by defau... | 0001_085_1085629_qa_2 | 6 |
[
{
"role": "system",
"content": "You are opencode, an interactive CLI tool that helps users with software engineering tasks. Use the instructions below and the tools available to you to assist the user.\n\nIMPORTANT: You must NEVER generate or guess URLs for the user unless you are confident that the URLs ar... | [
{
"type": "function",
"function": {
"name": "bash",
"description": "Executes a given bash command in a persistent shell session with optional timeout, ensuring proper handling and security measures.\n\nBe aware: OS: linux, Shell: bash\n\nAll commands run in the current working directory by defau... | 0001_137_1137361_qa_2 | 5 |
[
{
"role": "system",
"content": "You are opencode, an interactive CLI tool that helps users with software engineering tasks. Use the instructions below and the tools available to you to assist the user.\n\nIMPORTANT: You must NEVER generate or guess URLs for the user unless you are confident that the URLs ar... | [
{
"type": "function",
"function": {
"name": "bash",
"description": "Executes a given bash command in a persistent shell session with optional timeout, ensuring proper handling and security measures.\n\nBe aware: OS: linux, Shell: bash\n\nAll commands run in the current working directory by defau... | 0001_137_1137537_qa_2 | 5 |
[
{
"role": "system",
"content": "You are opencode, an interactive CLI tool that helps users with software engineering tasks. Use the instructions below and the tools available to you to assist the user.\n\nIMPORTANT: You must NEVER generate or guess URLs for the user unless you are confident that the URLs ar... | [
{
"type": "function",
"function": {
"name": "bash",
"description": "Executes a given bash command in a persistent shell session with optional timeout, ensuring proper handling and security measures.\n\nBe aware: OS: linux, Shell: bash\n\nAll commands run in the current working directory by defau... | 0001_155_1155051_qa_5 | 6 |
[
{
"role": "system",
"content": "You are opencode, an interactive CLI tool that helps users with software engineering tasks. Use the instructions below and the tools available to you to assist the user.\n\nIMPORTANT: You must NEVER generate or guess URLs for the user unless you are confident that the URLs ar... | [
{
"type": "function",
"function": {
"name": "bash",
"description": "Executes a given bash command in a persistent shell session with optional timeout, ensuring proper handling and security measures.\n\nBe aware: OS: linux, Shell: bash\n\nAll commands run in the current working directory by defau... | 0001_155_1155264_qa_5 | 6 |
[{"role":"system","content":"You are opencode, an interactive CLI tool that helps users with softwar(...TRUNCATED) | [{"type":"function","function":{"name":"bash","description":"Executes a given bash command in a pers(...TRUNCATED) | 0001_160_1160639_qa_1 | 6 |
[{"role":"system","content":"You are opencode, an interactive CLI tool that helps users with softwar(...TRUNCATED) | [{"type":"function","function":{"name":"bash","description":"Executes a given bash command in a pers(...TRUNCATED) | 0001_170_1170198_qa_3 | 7 |
[{"role":"system","content":"You are opencode, an interactive CLI tool that helps users with softwar(...TRUNCATED) | [{"type":"function","function":{"name":"bash","description":"Executes a given bash command in a pers(...TRUNCATED) | 0001_170_1170198_qa_4 | 6 |
[{"role":"system","content":"You are opencode, an interactive CLI tool that helps users with softwar(...TRUNCATED) | [{"type":"function","function":{"name":"bash","description":"Executes a given bash command in a pers(...TRUNCATED) | 0001_173_1173665_qa_3 | 5 |
[{"role":"system","content":"You are opencode, an interactive CLI tool that helps users with softwar(...TRUNCATED) | [{"type":"function","function":{"name":"bash","description":"Executes a given bash command in a pers(...TRUNCATED) | 0001_173_1173665_qa_5 | 5 |
End of preview. Expand in Data Studio
SmolDataEnvs OpenCode traces
Conversations of Qwen3.5-397B-A17B working as an
OpenCode agent through SmolDataEnvs
data-analysis tasks, one row per task it solved, ready for TRL's SFTTrainer.
from datasets import load_dataset
dataset = load_dataset("sergiopaniego/smoldataenvs-opencode-traces", "full", split="train")
| Config | Rows | Used by |
|---|---|---|
500 |
497 (473 trained on) | Qwen3.5-2B-SmolDataEnvs-SFT-500 |
1k |
1,000, the 497 plus 503 sampled (963 trained on) | Qwen3.5-2B-SmolDataEnvs-SFT-1k |
full |
2,307, every solved clean training task (2,238 trained on) | Qwen3.5-2B-SmolDataEnvs-SFT-full |
Columns
messages: the agent's last request to the model (system prompt, task, every assistant turn with its tool calls, every tool output) plus the final reply. Tool-call arguments are dicts.tools: the tools OpenCode offered, in OpenAI format.task_name,n_turns: the SmolDataEnvs task and the number of model calls the rollout took.
messages and tools are typed Json, so each tool call keeps only its own arguments.
How it was made
OpenEnv's Harbor integration ran OpenCode on each training task in a Hugging Face Sandbox, with the teacher served by
Inference Providers, and recorded every model call. The task's verifier graded the answer, and only rollouts scored 1
became rows. The teacher solved 2,307 of the 2,879 clean training tasks (80%), in a median of 6 turns. "Trained on"
counts the rows that fit in 16k tokens: sft.py drops longer ones rather than truncating their final answer. The teacher's reasoning is not in
the rows.
Only training tasks whose Kaggle dataset backs no task of the SmolDataEnvs eval split were used: 2,121 of the 5,000
training tasks share a dataset with eval, and a model trained on them has already seen the eval tasks' files. The
scripts are in TRL's examples/sft_agent_traces, and every raw rollout is in
sergiopaniego/smoldataenvs-opencode-rollouts.
License
MIT, as SmolDataEnvs. The conversations were generated by Qwen3.5-397B-A17B (Apache-2.0).
- Downloads last month
- -
Models trained or fine-tuned on sergiopaniego/smoldataenvs-opencode-traces