prompt listlengths 2 16 | ground_truth stringlengths 1k 2.87k | raw_prompt listlengths 2 16 | extra_info dict | reward_model dict | data_source stringclasses 1
value | agent_name stringclasses 1
value | index stringlengths 36 36 |
|---|---|---|---|---|---|---|---|
[
{
"content": "The assistant is designed to be helpful, proactive, and highly interactive.\n\nThe assistant strives to accurately interpret the user's intent throughout the conversation, acknowledging previous\ninteractions to maintain context and continuity. If the user's message is unclear or lacks necessary d... | Photo by Comeup 2021 on Comeup 2021
The global startup festival “COMEUP 2021” begins its three-day schedule event from its opening ceremony on the 17th until the 19th. Under the theme of Meet the Future — Transformation, “COMEUP 2021” presents ways for startups to view changes in the market environment and prepare for... | [
{
"content": "The assistant is designed to be helpful, proactive, and highly interactive.\n\nThe assistant strives to accurately interpret the user's intent throughout the conversation, acknowledging previous\ninteractions to maintain context and continuity. If the user's message is unclear or lacks necessary d... | {
"completion": "While I don't have direct access to specific transcripts or recordings from COMEUP 2021, I can suggest some approaches and sources to help you find more detailed insights about Lee Ju-hwan's session, particularly regarding communication costs in remote work.\n\n### Suggested Approaches:\n\n1. **Offic... | {
"ground_truth": "Photo by Comeup 2021 on Comeup 2021\n\nThe global startup festival “COMEUP 2021” begins its three-day schedule event from its opening ceremony on the 17th until the 19th. Under the theme of Meet the Future — Transformation, “COMEUP 2021” presents ways for startups to view changes in the market envi... | collabllm | collabllm_agent | a2f697eb-c038-4134-9e84-675bd3103b9f |
[
{
"content": "The assistant is designed to be helpful, proactive, and highly interactive.\n\nThe assistant strives to accurately interpret the user's intent throughout the conversation, acknowledging previous\ninteractions to maintain context and continuity. If the user's message is unclear or lacks necessary d... | Filter on the name
Filtering on the name goes via a controller; the type of filter inherently part of the table view. You only see what matches. To get a match we use a column containing the full name (thus first name and last name put together). The function Lower() is applied to flatten out capitals and the function... | [
{
"content": "The assistant is designed to be helpful, proactive, and highly interactive.\n\nThe assistant strives to accurately interpret the user's intent throughout the conversation, acknowledging previous\ninteractions to maintain context and continuity. If the user's message is unclear or lacks necessary d... | {
"completion": "Certainly! Let's address each of your questions in detail:\n\n### Discrepancy Between `!= -1` and `> 0`\n\n1. **`!= -1` Explanation:**\n - **Purpose:** Checks if the substring exists anywhere in the string, including the very beginning (index 0).\n - **Use Case:** Ideal for general searches where... | {
"ground_truth": "Filter on the name\n\nFiltering on the name goes via a controller; the type of filter inherently part of the table view. You only see what matches. To get a match we use a column containing the full name (thus first name and last name put together). The function Lower() is applied to flatten out ca... | collabllm | collabllm_agent | ff62a326-5fec-4d06-8923-f2c0239d8268 |
[
{
"content": "The assistant is designed to be helpful, proactive, and highly interactive.\n\nThe assistant strives to accurately interpret the user's intent throughout the conversation, acknowledging previous\ninteractions to maintain context and continuity. If the user's message is unclear or lacks necessary d... | Facebook is widely used in the United States. However, you can also use it to further your business. You can pay to do marketing on Facebook, but there are free options also. Whether you use the free options or the paid options, the following advice will help.
Use all of the options available to you on Facebook when u... | [
{
"content": "The assistant is designed to be helpful, proactive, and highly interactive.\n\nThe assistant strives to accurately interpret the user's intent throughout the conversation, acknowledging previous\ninteractions to maintain context and continuity. If the user's message is unclear or lacks necessary d... | {
"completion": "Certainly! Increasing likes and shares is a great goal as it can significantly extend your reach and boost engagement. Here’s how you can connect the strategies to achieving more likes and shares:\n\n### Increasing Likes and Shares\n\n1. **Optimize Post Timing**: Analyze when your audience is most ac... | {
"ground_truth": "Facebook is widely used in the United States. However, you can also use it to further your business. You can pay to do marketing on Facebook, but there are free options also. Whether you use the free options or the paid options, the following advice will help.\n\nUse all of the options available to... | collabllm | collabllm_agent | 934033cd-c654-4edd-8e05-96f00800c54c |
[
{
"content": "The assistant is designed to be helpful, proactive, and highly interactive.\n\nThe assistant strives to accurately interpret the user's intent throughout the conversation, acknowledging previous\ninteractions to maintain context and continuity. If the user's message is unclear or lacks necessary d... | With new rules coming into force from 2025 for rental properties, landlords must make sure their property’s EPC rating is C or above, or risk not being able to take on new tenants.
As a result, around 17% of landlords have made a conscious effort to improve the energy efficiency of their property. With a further 23% o... | [
{
"content": "The assistant is designed to be helpful, proactive, and highly interactive.\n\nThe assistant strives to accurately interpret the user's intent throughout the conversation, acknowledging previous\ninteractions to maintain context and continuity. If the user's message is unclear or lacks necessary d... | {
"completion": "To include specific statistics and examples in your article, you could look into recent surveys or reports by organizations like the National Residential Landlords Association (NRLA) or similar bodies, which might provide insights like '1 in 5 landlords are making EPC improvements.' Additionally, gov... | {
"ground_truth": "With new rules coming into force from 2025 for rental properties, landlords must make sure their property’s EPC rating is C or above, or risk not being able to take on new tenants.\n\nAs a result, around 17% of landlords have made a conscious effort to improve the energy efficiency of their propert... | collabllm | collabllm_agent | a7c08bbe-7975-4867-bacf-55a6434e8e12 |
[{"content":"The assistant is designed to be helpful, proactive, and highly interactive.\n\nThe assi(...TRUNCATED) | "Download luxury Nissan GTR Wallpaper for your desktop and mobile screens. These high-resolution ima(...TRUNCATED) | [{"content":"The assistant is designed to be helpful, proactive, and highly interactive.\n\nThe assi(...TRUNCATED) | {"completion":"To find a specific iconic Nissan GTR wallpaper that is highly rated among fans, I rec(...TRUNCATED) | {"ground_truth":"Download luxury Nissan GTR Wallpaper for your desktop and mobile screens. These hig(...TRUNCATED) | collabllm | collabllm_agent | fb6c241d-fab2-4ed5-b580-fa3d6f5468e5 |
[{"content":"The assistant is designed to be helpful, proactive, and highly interactive.\n\nThe assi(...TRUNCATED) | "Samaritan to the car guards in South Africa\n\nCar guards in South Africa\n\nIf you’re a South Af(...TRUNCATED) | [{"content":"The assistant is designed to be helpful, proactive, and highly interactive.\n\nThe assi(...TRUNCATED) | {"completion":"To create a comprehensive short article about car guards in South Africa, here's an o(...TRUNCATED) | {"ground_truth":"Samaritan to the car guards in South Africa\n\nCar guards in South Africa\n\nIf you(...TRUNCATED) | collabllm | collabllm_agent | fb81f972-b17a-4087-af7a-48de80e05f30 |
[{"content":"The assistant is designed to be helpful, proactive, and highly interactive.\n\nThe assi(...TRUNCATED) | "As titled said, I want to help you explore more in depth useful of NFT. As we already know NFT is u(...TRUNCATED) | [{"content":"The assistant is designed to be helpful, proactive, and highly interactive.\n\nThe assi(...TRUNCATED) | {"completion":"Incorporating testimonials into NFT resumes can be a powerful way to add credibility (...TRUNCATED) | {"ground_truth":"As titled said, I want to help you explore more in depth useful of NFT. As we alrea(...TRUNCATED) | collabllm | collabllm_agent | 93f25eae-d5e6-40f2-a3a3-f5a28f7a1727 |
[{"content":"The assistant is designed to be helpful, proactive, and highly interactive.\n\nThe assi(...TRUNCATED) | "I was talking about cake when a friend texted back saying that he loves Tres Leches. I was looking (...TRUNCATED) | [{"content":"The assistant is designed to be helpful, proactive, and highly interactive.\n\nThe assi(...TRUNCATED) | {"completion":"Let's delve a bit deeper into the linguistic side of your story! Here's an updated ve(...TRUNCATED) | {"ground_truth":"I was talking about cake when a friend texted back saying that he loves Tres Leches(...TRUNCATED) | collabllm | collabllm_agent | a73ee2cc-5bcd-438f-83cf-ab093fe70fb0 |
[{"content":"The assistant is designed to be helpful, proactive, and highly interactive.\n\nThe assi(...TRUNCATED) | "As one of the newer diets on the market, the Zone Diet is rapidly taking hold as one of the best an(...TRUNCATED) | [{"content":"The assistant is designed to be helpful, proactive, and highly interactive.\n\nThe assi(...TRUNCATED) | {"completion":"Certainly! Let's expand on the section about how the Zone Diet works by incorporating(...TRUNCATED) | {"ground_truth":"As one of the newer diets on the market, the Zone Diet is rapidly taking hold as on(...TRUNCATED) | collabllm | collabllm_agent | 00ce974e-f69a-40f5-9b58-c407f552de5a |
[{"content":"The assistant is designed to be helpful, proactive, and highly interactive.\n\nThe assi(...TRUNCATED) | "Galaxy Samsung Galaxy S21 FE Full Phone Specification: Everything You Need to Know\n\nin this blog (...TRUNCATED) | [{"content":"The assistant is designed to be helpful, proactive, and highly interactive.\n\nThe assi(...TRUNCATED) | {"completion":"It's understandable to encounter conflicting information, especially with variations (...TRUNCATED) | {"ground_truth":"Galaxy Samsung Galaxy S21 FE Full Phone Specification: Everything You Need to Know\(...TRUNCATED) | collabllm | collabllm_agent | 33fd89b9-4332-46fa-8dfb-ecbffa26e300 |
CollabLLM medium — RL (GRPO) train/validation split
Inputs for GRPO training on the CollabLLM medium document-writing task, as used to produce
the verl-grpo-medium-qwen3-4b-step{50,100,129} checkpoints.
| file | rows | size |
|---|---|---|
rl_train.parquet |
2072 | 16M |
rl_validation.parquet |
— | 2.2M |
Derived from the CollabLLM medium task (source articles: Kamaljp/medium_articles). Each row is
a single-turn prompt that the training loop expands into a multi-turn conversation with a
simulated user, then scores with an LLM judge.
⚠️ Known issue: train/eval overlap causes label leakage through the reward
Measured, not suspected. 35 of the 100 medium test prompts also appear in rl_train.parquet,
and on those rows the BLEU reward target is byte-identical to the test reference document.
Why this matters more for RL than for SFT: in supervised fine-tuning the reference never enters
the training text, so an overlap is only topical leakage. In RL the reference is the reward
target for the bleu_score metric, so an overlapping row leaks the eval label directly through
the reward signal.
Root cause: the upstream release sampled the example pool before applying its own train/test split, so the split does not separate them.
If you train on this, hold out the overlapping prompts, or report numbers on the non-overlapping 65. The overlap is a property of this file and travels with it.
Reproducing
The reward is a weighted sum of bleu_score (+1), interactivity (+1) and token_amount
(−0.1), with the judge and user simulator served as separate vLLM endpoints. Reference three-role
setting, thinking off everywhere:
| role | model |
|---|---|
| assistant (trained) | Qwen/Qwen3-4B |
| user simulator | Qwen/Qwen3.5-9B |
| judge / reward | Qwen/Qwen3.5-27B |
Full setup and the acceptance tests (5-step training replay, val@0 = 1.0618) are in
studynotes/new_cluster_onboarding.md in the code repo.
Note on the checkpoints
Training peaked at step 100 and regressed afterwards: step 129 has flat BLEU and
significantly worse interactivity. Report step 100. Do not pick a checkpoint on verl's
critic/rewards/mean — it credits brevity, which the benchmark judge penalises.
- Downloads last month
- 28