Cloud models just can't offer some features by design: privacy, low-latency, offline connectivity, free generation, etc. Not every application requires 2T parameters. Probably most don't.
@maximelabonne
-
Co-developing AI Architectures and Models for Customer Use Cases
By
–
In a sense, yes, we can co-develop architectures and models for specific use cases with customers.
-
Small Reasoning Models Training Research Artifact Released
By
–
Thanks to my excellent colleagues @timseyde and @rohin_manvi
, who did most of this work! We hope this research artifact will be useful as a recipe for training small reasoning models. More to come very soon! 🙂 -
SFT and GRPO Improve LFM-1B Math Reasoning Performance
By
–
The most interesting part is that the base LFM-1B model isn't that strong in math (see results). Extensive SFT (~100B tokens) was enough to turn it into a strong reasoner. Further GRPO compressed the reasoning traces and even maintained performance.
-
Fine-tuned Model Response Length Analysis Across AI Models
By
–
You can see the mean response length for correct and incorrect answers across different models. The source model is on the left and the fine-tuned model on the right. Average performance of the fine-tuned model is reported at the top as "32k score | 4k score".
-
GRPO 1.3B Model Achieves Best Performance Edge Applications
By
–
For edge applications (it's a 1B model!), this is not enough. We need more concise traces. With a 4k token budget, the GRPO (LFM-1.3B-Math) model achieves best-in-class performance. Note that it only cost us a few points on average at the 32k token setting.
-
LFM-1.3B-Distill SFT Model Outperforms Larger DeepSeek Competitor
By
–
Evals show that our SFT (LFM-1.3B-Distill) model performs slightly better with a 32k token budget. It's actually competitive with models based on DeepSeek-R1-Distill-Qwen-1.5B while being 15% smaller.
-
GRPO with Difficulty-Aware Rewards for AI Model Training
By
–
We leveraged GRPO with a difficulty-aware reward formulation to address this issue. (More information about our custom GRPO flavor in the article.) We combined it with the following data mix after filtering out samples that do not yield a solve rate of 20-80% at 4k tokens.
-
SFT Fine-tuning Strategy for 1.3B Chat Model
By
–
The first step is SFTing our general-purpose 1.3B chat model on the right mixture. We iterated over many open-source datasets to find the ideal mix. This is great to raise performance, but also makes the model extremely verbose (>10k tokens on average).
-
Small 1B Model Achieves Strong Reasoning Performance
By
–
Can small models be strong reasoners? We created a 1B reasoning model at @liquidai that is both accurate and concise We applied a combination of SFT (to raise quality) and GRPO (to control verbosity) The result is a best-in-class model without specific math pre-training