DaisyRose-27B: 1,200-Task RL Boosts Qwen3.8 from 5.6% to 12% on Terminal-Bench 4.1

DaisyRose-27B uses lightweight RL on 1,200 tasks to double Qwen3.8's terminal benchmark score from 5.6% to 12%.
DaisyRose-27B shows that lightweight RL can significantly improve terminal operation tasks: with only 1,200 training tasks, Qwen3.8's Terminal-Bench 4.1 score rose from 5.6% to 12%, more than doubling. Clear, auto-verifiable feedback makes reward design efficient, and task quality matters more than quantity. A 27B model with fine-grained RL offers a cost-controllable path for small teams.
A Lightweight RL Experiment Delivers a Striking Leap
The DaisyRose-27B project, which sparked discussion on Hacker News, demonstrates a noteworthy reinforcement learning (RL) experiment: after training on 1,200 tasks, Qwen3.8's score on the Terminal-Bench 4.1 benchmark rose from 5.6% to 12%. The number itself isn't high, but a relative improvement of more than double shows that targeted RL training can indeed bring substantive change to complex tasks like terminal operations.
Terminal-Bench 4.1 is a benchmark that evaluates a model's ability to complete practical tasks in command-line environments, covering scenarios such as file operations, script execution, and environment configuration. Qwen3.8's original 5.6% score means it performed very poorly on such tasks, and while 12% remains low, an improvement of more than double is a considerable result for a small-scale RL experiment.
Why 1,200 Tasks Are Enough to Make a Difference
The most intriguing aspect of this experiment is the training scale. 1,200 tasks is a relatively lightweight amount of data for RL training, far short of large-scale training that involves tens of thousands or even hundreds of thousands of samples. But DaisyRose-27B's results suggest that task quality and specificity may matter more than quantity.
Terminal tasks have a distinctive characteristic: feedback signals are relatively clear. Whether a command executes successfully, whether a file is created correctly, whether output matches expectations — all of these can be verified automatically. This verifiability makes reward design for RL straightforward, allowing the model to obtain clear success/failure signals from each attempt. If 1,200 tasks cover a sufficiently diverse range of terminal operation scenarios, they are enough for the model to learn some general behavioral patterns.
The jump from 5.6% to 12% also implies that Qwen3.8's baseline capabilities are not entirely absent — it has some foundation in instruction understanding and code generation, and RL training is more about activating and guiding these existing capabilities rather than building them from scratch.
When training large language models with reinforcement learning (RL), reward signal design is the core difficulty. In dialogue or open-ended generation tasks, judging whether a response is "good" often requires human annotation or training an additional reward model, which is costly and noisy. Terminal tasks, by contrast, naturally lend themselves to programmatic verification: command exit codes, filesystem state changes, and pattern matching on standard output can all be used to automatically determine success or failure via scripts. This "verifiable reward" makes RL's feedback loop cheap and reliable — the model receives a definite yes/no signal with every attempt, enabling more efficient policy updates. This is also why in recent years, RL training has generally outperformed in domains with objective answers, such as mathematics and code, compared with open-ended tasks.
The Pragmatic Path of Small Models + RL
The "27B" in DaisyRose-27B's name suggests this may be a model solution at the 27-billion-parameter level. Amid the current trend of pursuing ever-larger parameter counts, this experiment offers an alternative approach: rather than continually stacking up model scale, it may be better to do fine-grained RL tuning on a medium-scale model.
From an engineering perspective, the advantage of this path is controllable cost. RL training on 1,200 tasks does not require massive compute, and small teams or even individual developers could potentially reproduce and iterate on it. The discussion heat on Hacker News — 34 points and 4 comments — isn't particularly high, but the value of such experiments often lies in reproducibility and scalability: if 1,200 tasks can bring this kind of improvement, then further optimizing task design and increasing task diversity may leave room for continued gains.
What needs to be viewed soberly is that the absolute level of 12% remains low. The tasks in Terminal-Bench 4.1 place high demands on a model's planning ability, error recovery ability, and familiarity with command-line tools. RL training can improve behavioral policies, but the model's basic world knowledge and reasoning ability remain bottlenecks.
27B-parameter models have an often-overlooked advantage in RL training: lower exploration cost. RL relies on the model generating large numbers of attempt trajectories (rollouts) during training. The larger the parameter count, the higher the compute and memory overhead per forward inference, and the time and cost required to collect enough samples rise proportionally. A medium-scale model can generate more trajectories and perform more rounds of policy updates under the same budget, which is especially critical for lightweight RL experiments that are inherently sensitive to sample efficiency. In addition, models at the 27B level can typically be fine-tuned on a single machine with multiple GPUs, or even on a single GPU after quantization, lowering the reproduction barrier and making "small teams doing RL" feasible rather than merely hypothetical.
Implications for Terminal Agent Development
This experiment has direct reference value for developers working on terminal agents. The terminal environment is a touchstone for testing a model's practical hands-on ability — unlike dialogue tasks, it cannot be faked through language fluency; every step of operation has an objective result.
DaisyRose-27B's results show that using verifiable tasks for RL is an effective path to improving a model's terminal operation capabilities. Developers can consider building their own domain-specific task sets — not necessarily large, but with clear success/failure criteria — and then using RL to elevate the model's performance in specific scenarios. This approach can be transferred to other scenarios requiring precise operations, such as database management, system operations and maintenance, and CI/CD pipeline operations.
Of course, experimental results from a single source still require more independent verification. Details such as the specific composition of the 1,200 tasks, the design of the reward function, and training hyperparameters will all affect the generalizability of the conclusions. But directionally, lightweight RL activating the potential of medium-scale models on specific tasks is a technical path worth watching.
Terminal agents typically adopt an "observe—think—act" loop architecture: the model reads the current terminal output, decides on the next command to execute, and then adjusts subsequent actions based on the execution result. The capability bottleneck of such agents often lies not in language understanding, but in multi-step planning and error recovery — whether it can diagnose the cause after a command fails and switch to the correct tool. RL training is effective precisely because it optimizes the cumulative return of the entire action sequence rather than the likelihood of a single-step output. When developers build their own task sets, they should prioritize designing scenarios that include branching decisions and failure retries, so that the model learns to recover from errors during training. This improves robustness in real environments more than simply increasing repetitions of easy tasks.
One-line Summary
DaisyRose-27B uses lightweight RL on 1,200 tasks to double Qwen3.8's terminal benchmark score from 5.6% to 12%.
Paragraph Summary
The DaisyRose-27B project demonstrates the striking effect of lightweight reinforcement learning on terminal operation tasks: trained on only 1,200 tasks, Qwen3.8's score on Terminal-Bench 4.1 rose from 5.6% to 12%, a relative improvement of more than double. The article argues that terminal tasks have clear, automatically verifiable feedback signals, making reward design direct and efficient, and that task quality and specificity matter more than quantity. A 27B medium-scale model combined with fine-grained RL tuning offers a pragmatic, cost-controllable path for resource-constrained teams. It also notes that the absolute level of 12% remains low, that the model's basic planning and reasoning capabilities remain bottlenecks, and that a single experimental result still requires independent verification.
Related articles

Inside Honey Trap Training: Emotional Dissociation and Anchoring Techniques
How are spies trained for honey trap missions? A trainee reveals emotional dissociation, perfume and keyword anchoring, and target photo binding—psychological techniques that condition reflexes for intelligence work.

Frontier AI Risk Management: Full-Pipeline Practices from Training to Deployment
OpenAI outlines a frontier AI risk management framework: from training-stage model alignment to deployment-stage sandboxing and blocking safeguards, plus post-deployment monitoring—a layered risk control system across the model lifecycle.

Show HN: Having Fun with Bicycle Trainers and Web APIs
A Show HN project on Hacker News, Fun with bicycle trainers and web APIs, explores combining bicycle trainers with Web APIs. This article analyzes its possible technical directions and significance.