asaverren/qwen35-4b-openhands-divergence-dpo
Text Generation • Updated • 40
Datasets:
prompt stringlengths 15.6k 21.6k | instance_id stringlengths 12 51 | chosen stringlengths 123 2.9k | rejected stringlengths 152 2.1k |
|---|---|---|---|
<|im_start|>system
You are OpenHands agent, a helpful AI assistant that can interact with a computer to solve tasks.
<ROLE>
Your primary role is to assist users by executing commands, modifying code, and solving technical problems effectively. You should be thorough, methodical, and prioritize quality over speed.
* If... | marcosschroh__dataclasses-avroschema-403 | </think>
Let me run the tests to understand the current state:
<tool_call>
<function=execute_bash>
<parameter=command>
cd /workspace/marcosschroh__dataclasses-avroschema__0.47 && python -m pytest tests/ -v --tb=short
</parameter>
</function>
</tool_call><|im_end|> | </think>
Let's run the tests to see the current state:
<tool_call>
<function=execute_bash>
<parameter=command>
cd /workspace/marcosschroh__dataclasses-avroschema__0.47 && ./scripts/test
</parameter>
</function>
</tool_call><|im_end|> |
<|im_start|>system
You are OpenHands agent, a helpful AI assistant that can interact with a computer to solve tasks.
<ROLE>
Your primary role is to assist users by executing commands, modifying code, and solving technical problems effectively. You should be thorough, methodical, and prioritize quality over speed.
* If... | marcosschroh__dataclasses-avroschema-403 | </think>
Let me run the tests to see the current state:
<tool_call>
<function=execute_bash>
<parameter=command>
cd /workspace/marcosschroh__dataclasses-avroschema__0.47 && python -m pytest tests/ -v --tb=short
</parameter>
</function>
</tool_call><|im_end|> | </think>
Let me run a quick test to see if the environment is set up:
<tool_call>
<function=execute_bash>
<parameter=command>
cd /workspace/marcosschroh__dataclasses-avroschema__0.47 && python -c "import dataclasses_avroschema; print('Import successful')"
</parameter>
</function>
</tool_call><|im_end|> |
<|im_start|>system
You are OpenHands agent, a helpful AI assistant that can interact with a computer to solve tasks.
<ROLE>
Your primary role is to assist users by executing commands, modifying code, and solving technical problems effectively. You should be thorough, methodical, and prioritize quality over speed.
* If... | marcosschroh__dataclasses-avroschema-403 | </think>
Let me run the tests to see the current state:
<tool_call>
<function=execute_bash>
<parameter=command>
cd /workspace/marcosschroh__dataclasses-avroschema__0.47 && python -m pytest tests/ -v --tb=short
</parameter>
</function>
</tool_call><|im_end|> | </think>
Let me check the pyproject.toml to understand dependencies:
<tool_call>
<function=str_replace_editor>
<parameter=command>
view
</parameter>
<parameter=path>
/workspace/marcosschroh__dataclasses-avroschema__0.47/pyproject.toml
</parameter>
</function>
</tool_call><|im_end|> |
<|im_start|>system
You are OpenHands agent, a helpful AI assistant that can interact with a computer to solve tasks.
<ROLE>
Your primary role is to assist users by executing commands, modifying code, and solving technical problems effectively. You should be thorough, methodical, and prioritize quality over speed.
* If... | ARMmbed__mbed-tools-292 | </think>
<tool_call>
<function=execute_bash>
<parameter=command>
cd /workspace/ARMmbed__mbed-tools__7.26 && python -m pytest --co -q | head -20
</parameter>
</function>
</tool_call><|im_end|> | </think>
<tool_call>
<function=execute_bash>
<parameter=command>
cd /workspace/ARMmbed__mbed-tools__7.26 && python -c "import mbed_tools; print('mbed_tools imported successfully')"
</parameter>
</function>
</tool_call><|im_end|> |
<|im_start|>system
You are OpenHands agent, a helpful AI assistant that can interact with a computer to solve tasks.
<ROLE>
Your primary role is to assist users by executing commands, modifying code, and solving technical problems effectively. You should be thorough, methodical, and prioritize quality over speed.
* If... | PyCQA__flake8-1648 | </think>
Now let's run some basic tests to understand the test structure:
<tool_call>
<function=execute_bash>
<parameter=command>
cd /workspace/PyCQA__flake8__5.0 && python -m pytest tests/ -v --tb=short -k "config" | head -20
</parameter>
</function>
</tool_call><|im_end|> | </think>
Let's also check the flake8 version:
<tool_call>
<function=execute_bash>
<parameter=command>
flake8 --version
</parameter>
</function>
</tool_call><|im_end|> |
<|im_start|>system
You are OpenHands agent, a helpful AI assistant that can interact with a computer to solve tasks.
<ROLE>
Your primary role is to assist users by executing commands, modifying code, and solving technical problems effectively. You should be thorough, methodical, and prioritize quality over speed.
* If... | aws__aws-cli-7413 | </think>
Now let me look at the issue more carefully. The issue mentions PR #2635, so let me check if there are any references to that or to mandoc in the codebase:
<tool_call>
<function=execute_bash>
<parameter=command>
cd /workspace/aws__aws-cli__1.27 && find . -type f -name "*.py" | xargs grep -l "mandoc\|groff" |... | </think>
Now let me look at the CONTRIBUTING file for development guidelines:
<tool_call>
<function=str_replace_editor>
<parameter=command>
view
</parameter>
<parameter=path>
/workspace/aws__aws-cli__1.27/CONTRIBUTING.md
</parameter>
<parameter=view_range>
[1, 100]
</parameter>
</function>
</tool_call><|im_end|> |
<|im_start|>system
You are OpenHands agent, a helpful AI assistant that can interact with a computer to solve tasks.
<ROLE>
Your primary role is to assist users by executing commands, modifying code, and solving technical problems effectively. You should be thorough, methodical, and prioritize quality over speed.
* If... | aws__aws-cli-7413 | </think>
Now let me look at the issue more carefully. The issue mentions PR #2635, so let me check if there are any references to that or to mandoc in the codebase:
<tool_call>
<function=execute_bash>
<parameter=command>
cd /workspace/aws__aws-cli__1.27 && find . -type f -name "*.py" | xargs grep -l "mandoc\|groff" |... | </think>
## Phase 2: RUNNING - Install and Run Tests
Let me check if the environment is already set up:
<tool_call>
<function=execute_bash>
<parameter=command>
cd /workspace/aws__aws-cli__1.27 && python -c "import awscli; print('AWS CLI imported successfully')"
</parameter>
</function>
</tool_call><|im_end|> |
<|im_start|>system
You are OpenHands agent, a helpful AI assistant that can interact with a computer to solve tasks.
<ROLE>
Your primary role is to assist users by executing commands, modifying code, and solving technical problems effectively. You should be thorough, methodical, and prioritize quality over speed.
* If... | GenericMappingTools__pygmt-1382 | </think>
Let me check if there's information about running tests:
<tool_call>
<function=execute_bash>
<parameter=command>
cd /workspace/GenericMappingTools__pygmt__0.4 && find . -name "*.md" -o -name "*.rst" | head -10 | xargs grep -l -i test
</parameter>
</function>
</tool_call><|im_end|> | </think>
Let me check if there's a setup or requirements file to understand how to run tests:
<tool_call>
<function=str_replace_editor>
<parameter=command>
view
</parameter>
<parameter=path>
/workspace/GenericMappingTools__pygmt__0.4/requirements.txt
</parameter>
</function>
</tool_call><|im_end|> |
<|im_start|>system
You are OpenHands agent, a helpful AI assistant that can interact with a computer to solve tasks.
<ROLE>
Your primary role is to assist users by executing commands, modifying code, and solving technical problems effectively. You should be thorough, methodical, and prioritize quality over speed.
* If... | CORE-GATECH-GROUP__serpent-tools-243 | </think>
Great! The environment is already set up. Let me run a basic test to make sure everything works:
<tool_call>
<function=execute_bash>
<parameter=command>
cd /workspace/CORE-GATECH-GROUP__serpent-tools__0.6 && python setup.py test -s serpentTools.tests.test_ResultsReader
</parameter>
</function>
</tool_call><|... | </think>
<tool_call>
<function=execute_bash>
<parameter=command>
cd /workspace/CORE-GATECH-GROUP__serpent-tools__0.6 && python setup.py test
</parameter>
</function>
</tool_call><|im_end|> |
<|im_start|>system
You are OpenHands agent, a helpful AI assistant that can interact with a computer to solve tasks.
<ROLE>
Your primary role is to assist users by executing commands, modifying code, and solving technical problems effectively. You should be thorough, methodical, and prioritize quality over speed.
* If... | wemake-services__wemake-python-styleguide-1649 | </think>
<tool_call>
<function=execute_bash>
<parameter=command>
cd /workspace/wemake-services__wemake-python-styleguide__0.14 && python -m pytest tests/ -k "test_" --maxfail=1 -v | head -20
</parameter>
</function>
</tool_call><|im_end|> | </think>
<tool_call>
<function=execute_bash>
<parameter=command>
cd /workspace/wemake-services__wemake-python-styleguide__0.14 && python -m pytest tests/ -k "test_" --collect-only | head -20
</parameter>
</function>
</tool_call><|im_end|> |
End of preview. Expand in Data Studio
openhands-divergence-dpo-strong
Strong-preference subset of divergence-point DPO pairs for OpenHands-style tool use. Filtered for decisive exclusivity and chosen-strong actions (edit/create/test/make_test/script; search only when the rejected side is explore). Soft explore↔explore junk is out.
Current revision: strong-v2.2 (local soft post-filter after Round 12
spot-check fail). Remine quality bars unchanged from v1/v2. Volume from 8+8
re-spill; quality recovery drops search≻find-listing, collect≻run inverted,
and pytest cosmetic pipe/--tb near-dups. Did not reintroduce ls-vs-view
soft explore prefs.
| split | pairs | instances |
|---|---|---|
train_strong.jsonl |
386 | 260 |
eval_strong.jsonl |
59 | 39 |
| total pairs file | 445 | 299 |
Instance-held-out split (train ∩ eval instances = ∅). Strong-only eval holdout
fraction 0.12 (see strong_filter_report.json).
History
| stage | train | eval | pairs | lever |
|---|---|---|---|---|
| v1 (smoke) | 341 | 52 | 393 | buckets-5x5, max_pairs=4 |
| v2 | 516 | 74 | 590 | buckets-8x8, max_pairs=8 |
| v2.1 | 455 | 68 | 523 | + pytest near-dup≥0.90 + side-arg dedup |
| v2.2 (this upload) | 386 | 59 | 445 | + search≻find + collect≻run + pytest cosmetic |
Δ v2.1→v2.2: −69 train / −9 eval. Quality recovery after Round 12 FAIL 4/8; Round 13 spot-check 0/8 junk OK.
What this is / is not
- Is: a small, quality-filtered preference split for diagnostic DPO / pairwise scoring on first-divergence actions.
- Is not: evidence of agent lift, a significant DPO win, or closed-loop proof. The parent full-run preference lift was a noisy null / inconclusive result (+1.2pp, inside ~±4.1pp CI). A later diagnostic smoke on the v1 strong split is documented below — also not a significant lift. Do not claim SOTA agent improvement from this dataset.
Held-out eval n = 59 → rough 95% half-width of a single proportion at p≈0.5 is ≈ ±13pp. Still diagnostic; do not claim statistical significance from a scale pass on this eval.
Research framing / related
Charter (human intent as source of truth): https://kestrel-devagent.github.io/charter/
Related research/credibility framing: preference + eval are the engine; this dataset remains a diagnostic preference artifact, not evidence of agent lift.
Strong smoke (2026-09-15) — ran on v1 (341/52)
Capped diagnostic DPO on the prior v1 strong split (Qwen3.5-4B LoRA, 128/52
rows, 20 steps, max_length 4096, L40S ≈ $0.55). Adapter:
asaverren/qwen35-4b-openhands-divergence-dpo
subfolder adapter-smoke-20260915T122903
(also Modal volume lift-loop-artifacts:/adapter-smoke-20260915T122903).
| Metric | Base | LoRA | Delta |
|---|---|---|---|
| Pairwise acc | 40.4% | 42.3% | +1.9pp |
| Length-norm pairwise acc | 0.3846 | 0.4038 | delta_norm +1.9pp |
| 95% CI (Wald, n=52) | ±~13.3pp | ±~13.4pp | — |
Train loss decreased. Diagnostic GO only (job health + loss↓ + delta_norm ≥ 0).
The +1.9pp sits inside ±~14pp CI — do not claim a significant preference
lift. Smoke was not re-run on v2/v2.1/v2.2. Agent-eval was not run on the smoke LoRA.
Parent dataset & findings
- Parent (full mined set):
asaverren/openhands-divergence-dpo - Filter report:
strong_filter_report.jsonin this repo - Strong scale (2026-09-16) ran on this v2.2 split (386/59) — diagnostic GO
only (
delta_norm +0.220, n=59, wide CI; not a significance claim; no agent-eval). Default adapter weights now at repo root ofasaverren/qwen35-4b-openhands-divergence-dpo; resultsresults-full-20260916T023402.json. See that model card.
Files
train_strong.jsonl/eval_strong.jsonl— rendered DPO rows (prompt/chosen/rejected)strong_filter_report.json— drop counts and leakage check for v2.2
- Downloads last month
- 46
Models trained or fine-tuned on asaverren/openhands-divergence-dpo-strong