Review Request: Trackio + RapidFire AI Co-Announcement Blog Post

[kamran-rapidfireAI]

Hey @qgallouedec! 👋

We'd love to get your review on our Trackio + RapidFire AI co-announcement blog post before publishing.

Article Location

Trackio-RapidFire-promotion/Co-Announcement Blog Trackio and RapidFire AI.md

Key Highlight

This article showcases Trackio being used for RAG experimentation for the first time — not just fine-tuning. We're tracking retrieval metrics like Precision, Recall, NDCG, and MRR alongside the typical training metrics.

What We're Looking For

• Accuracy of Trackio-related content
• Any suggestions or improvements
• Approval for the co-announcement

Thanks for your time!

[qgallouedec]

Nice, reading through the article, here are a few remarks:

---

• instead of the two figure, you can try playing with the leveraging an iframe, see https://huggingface.co/docs/trackio/deploy_embed#embedding-a-trackio-dashboard

Trackio is included as a dependency of RapidFire AI

so it's a required dep? cool move!

---

in the notebook, you have

def sample_formatting_function(example):
    return {
        "text": f"Question: {example['instruction']}\nAnswer: {example['response']}",
        "instruction": example['instruction'],  # Keep original
        "response": example['response']  # Keep original
    }

eval_dataset = eval_dataset.map(sample_formatting_function)

these two lines can be safely removed

        "instruction": example['instruction'],  # Keep original
        "response": example['response']  # Keep original

map will keep the unmodified column as they were

---

fp16=True,
...
"torch_dtype": "float16",  # Explicit fp16

FP16 is known to be less stable. We usually don't recommend it. I'd recommend using BF16 instead, or FP32.

---

model_name="gpt2",

for your examples I'd strongly recommend more recent models, like Qwen/Qwen3-0.6B or google/gemma-3-270m-it if you want even smaller

---

These are the same, no?

    if model_type == "causal_lm":
        model = AutoModelForCausalLM.from_pretrained(model_name, **model_kwargs)
    else:
        # Default to causal LM
        model = AutoModelForCausalLM.from_pretrained(model_name, **model_kwargs)

[qgallouedec]

Approval for the co-announcement

Ok for me :)

[kbigdelysh]

Thanks for the feedback, @qgallouedec! Really appreciate you taking the time to review.

Quick question — beyond publishing this blog post, do you have any suggestions on how we can increase awareness of this integration? We'd love to get RapidFire AI in front of more ML practitioners and teams, especially those already in the Hugging Face ecosystem.

A few things we're considering:

• Cross-posting on Hugging Face blog or community channels
• Social media announcements (Twitter/X, LinkedIn)
• Showcasing in TRL or Trackio documentation/examples

Given your experience growing adoption for TRL and other Hugging Face projects, we'd really value any insights you might have on what's worked well for reaching both individual developers and companies.

cc: @arun-rfai , @Js-norris

[kamran-rapidfireAI]

@qgallouedec Can you please publish the article next week? (we are almost finishing testing our fixes to the notebooks referred in the article). Thank you.

[qgallouedec]

Awesome! I think you should be able to do it yourself via your hugging face account?

[kamran-rapidfireAI]

@qgallouedec HuggingFace website does now show us a way to publish the blog post in the blog's main section. Instead it automatically goes to the community section of the blog (which we don't want because of low visibility). Is it possible you publish it as we are promoting Trackio? P.S. our organization has a Team membership but based on an HF article, "Articles published by Team or Enterprise organizations with 30+ seats will be featured in the main section of the Hugging Face Blog." We don't have 30 seats.

cc: @Js-norris @arun-rfai