VRAG-RL: Empower Vision-Perception-Based RAG for Visually Rich Information Understanding via Iterative Reasoning with Reinforcement Learning
Paper
•
2505.22019
•
Published
•
10
Top Contributors: Model Downloads
community
AI & ML interests
🏛️ Creators of models with the most cumulative new downloads each month (users only, no orgs)
Recent Activity
TopContributors-ModelDownloads's activity
Lin-Chen
authored
a
paper
9 days ago
Lin-Chen
authored
a
paper
25 days ago
eugenesiow
posted
an
update
about 2 months ago
Post
1568
GPT-4.1 dropped this week - and it puts OpenAI back in the race for coding & agentic leadership.
⚙️ API only - no ChatGPT toggle for this.
💻 Coding performance is back on par with Claude 3.7 Sonnet & Gemini 2.5 Pro (though Gemini still leads).
💸 Pricing:
• Full: $3.50 / 1M tokens
• Mini: $0.70 / 1M
• Nano: $0.17 / 1M
👉 Gemini 2.5 Pro = best price/perf ($3.44 / 1M)
😵 Claude 3.5 Sonnet = $6 / 1M (!)
🧠 Not a "thinking" model.
📊 Mini shines on general reasoning tasks (e.g. GPQA), but only the full model holds up in SWE-bench-verified (GitHub issue solving).
⚙️ API only - no ChatGPT toggle for this.
💻 Coding performance is back on par with Claude 3.7 Sonnet & Gemini 2.5 Pro (though Gemini still leads).
💸 Pricing:
• Full: $3.50 / 1M tokens
• Mini: $0.70 / 1M
• Nano: $0.17 / 1M
👉 Gemini 2.5 Pro = best price/perf ($3.44 / 1M)
😵 Claude 3.5 Sonnet = $6 / 1M (!)
🧠 Not a "thinking" model.
📊 Mini shines on general reasoning tasks (e.g. GPQA), but only the full model holds up in SWE-bench-verified (GitHub issue solving).
Post
32767
Access requests enabled for latest GLM models
While a fix is being implemented (https://github.com/ggml-org/llama.cpp/pull/12957) I want to leave the models up for visibility and continued discussion, but want to prevent accidental downloads of known broken models (even though there are settings that could fix it at runtime for now)
With this goal, I've enabled access requests. I don't really want your data, so I'm sorry that I don't think there's a way around that? But that's what I'm gonna do for now, and I'll remove the gate when a fix is up and verified and I have a chance to re-convert and quantize!
Hope you don't mind in the mean time :D
While a fix is being implemented (https://github.com/ggml-org/llama.cpp/pull/12957) I want to leave the models up for visibility and continued discussion, but want to prevent accidental downloads of known broken models (even though there are settings that could fix it at runtime for now)
With this goal, I've enabled access requests. I don't really want your data, so I'm sorry that I don't think there's a way around that? But that's what I'm gonna do for now, and I'll remove the gate when a fix is up and verified and I have a chance to re-convert and quantize!
Hope you don't mind in the mean time :D
- 1 reply
Lin-Chen
authored
a
paper
about 2 months ago
mvaloatto
updated
a
Space
4 months ago
Post
73200
Switching to
I posted a poll on twitter, and others have mentioned the interest in me using the convention of including the author name in the model path when I upload.
It has a couple advantages, first and foremost of course is ensuring clarity of who uploaded the original model (did Qwen upload Qwen2.6? Or did someone fine tune Qwen2.5 and named it 2.6 for fun?)
The second thing is that it avoids collisions, so if multiple people upload the same model and I try to quant them both, I would normally end up colliding and being unable to upload both
I'll be implementing the change next week, there are just two final details I'm unsure about:
First, should the files also inherit the author's name?
Second, what to do in the case that the author name + model name pushes us past the character limit?
Haven't yet decided how to handle either case, so feedback is welcome, but also just providing this as a "heads up"
author_model-nameI posted a poll on twitter, and others have mentioned the interest in me using the convention of including the author name in the model path when I upload.
It has a couple advantages, first and foremost of course is ensuring clarity of who uploaded the original model (did Qwen upload Qwen2.6? Or did someone fine tune Qwen2.5 and named it 2.6 for fun?)
The second thing is that it avoids collisions, so if multiple people upload the same model and I try to quant them both, I would normally end up colliding and being unable to upload both
I'll be implementing the change next week, there are just two final details I'm unsure about:
First, should the files also inherit the author's name?
Second, what to do in the case that the author name + model name pushes us past the character limit?
Haven't yet decided how to handle either case, so feedback is welcome, but also just providing this as a "heads up"
- 5 replies
Lin-Chen
authored
a
paper
6 months ago
Post
80392
Looks like Q4_0_N_M file types are going away
Before you panic, there's a new "preferred" method which is online (I prefer the term on-the-fly) repacking, so if you download Q4_0 and your setup can benefit from repacking the weights into interleaved rows (what Q4_0_4_4 was doing), it will do that automatically and give you similar performance (minor losses I think due to using intrinsics instead of assembly, but intrinsics are more maintainable)
You can see the reference PR here:
https://github.com/ggerganov/llama.cpp/pull/10446
So if you update your llama.cpp past that point, you won't be able to run Q4_0_4_4 (unless they add backwards compatibility back), but Q4_0 should be the same speeds (though it may currently be bugged on some platforms)
As such, I'll stop making those newer model formats soon, probably end of this week unless something changes, but you should be safe to download and Q4_0 quants and use those !
Also IQ4_NL supports repacking though not in as many shapes yet, but should get a respectable speed up on ARM chips, PR for that can be found here: https://github.com/ggerganov/llama.cpp/pull/10541
Remember, these are not meant for Apple silicon since those use the GPU and don't benefit from the repacking of weights
Before you panic, there's a new "preferred" method which is online (I prefer the term on-the-fly) repacking, so if you download Q4_0 and your setup can benefit from repacking the weights into interleaved rows (what Q4_0_4_4 was doing), it will do that automatically and give you similar performance (minor losses I think due to using intrinsics instead of assembly, but intrinsics are more maintainable)
You can see the reference PR here:
https://github.com/ggerganov/llama.cpp/pull/10446
So if you update your llama.cpp past that point, you won't be able to run Q4_0_4_4 (unless they add backwards compatibility back), but Q4_0 should be the same speeds (though it may currently be bugged on some platforms)
As such, I'll stop making those newer model formats soon, probably end of this week unless something changes, but you should be safe to download and Q4_0 quants and use those !
Also IQ4_NL supports repacking though not in as many shapes yet, but should get a respectable speed up on ARM chips, PR for that can be found here: https://github.com/ggerganov/llama.cpp/pull/10541
Remember, these are not meant for Apple silicon since those use the GPU and don't benefit from the repacking of weights
-
- 17 replies
Lin-Chen
authored
a
paper
6 months ago
Post
16494
Old mixtral model quants may be broken!
Recently Slaren over on llama.cpp refactored the model loader - in a way that's super awesome and very powerful - but with it came breaking of support for "split tensor MoE models", which applies to older mixtral models
You may have seen my upload of one such older mixtral model, ondurbin/bagel-dpo-8x7b-v0.2, and with the newest changes it seems to be able to run without issue
If you happen to run into issues with any other old mixtral models, drop a link here and I'll try to remake them with the new changes so that we can continue enjoying them :)
Recently Slaren over on llama.cpp refactored the model loader - in a way that's super awesome and very powerful - but with it came breaking of support for "split tensor MoE models", which applies to older mixtral models
You may have seen my upload of one such older mixtral model, ondurbin/bagel-dpo-8x7b-v0.2, and with the newest changes it seems to be able to run without issue
If you happen to run into issues with any other old mixtral models, drop a link here and I'll try to remake them with the new changes so that we can continue enjoying them :)
-
- 2 replies
Post
23924
In regards to the latest mistral model and GGUFs for it:
Yes, they may be subpar and may require changes to llama.cpp to support the interleaved sliding window
Yes, I got excited when a conversion worked and released them ASAP
That said, generation seems to work right now and seems to mimic the output from spaces that are running the original model
I have appended -TEST to the model names in an attempt to indicate that they are not final or perfect, but if people still feel mislead and that it's not the right thing to do, please post (civilly) below your thoughts, I will highly consider pulling the conversions if that's what people think is best. After all, that's what I'm here for, in service to you all !
Yes, they may be subpar and may require changes to llama.cpp to support the interleaved sliding window
Yes, I got excited when a conversion worked and released them ASAP
That said, generation seems to work right now and seems to mimic the output from spaces that are running the original model
I have appended -TEST to the model names in an attempt to indicate that they are not final or perfect, but if people still feel mislead and that it's not the right thing to do, please post (civilly) below your thoughts, I will highly consider pulling the conversions if that's what people think is best. After all, that's what I'm here for, in service to you all !
-
- 6 replies
Post
35026
Reposting from twitter:
Just so you all know, I'll be on vacation for the following two weeks and away from home! I'm hoping to get on at least once a day to load up some quants, but I won't be as bleeding edge and on the ball :) feel free to shoot me a message if you see one I should make!
In the meantime if you need something bleeding edge make sure to check out @MaziyarPanahi or @bullerwins who both put out great work!
Just so you all know, I'll be on vacation for the following two weeks and away from home! I'm hoping to get on at least once a day to load up some quants, but I won't be as bleeding edge and on the ball :) feel free to shoot me a message if you see one I should make!
In the meantime if you need something bleeding edge make sure to check out @MaziyarPanahi or @bullerwins who both put out great work!
- 4 replies
Post
16289
Decided to try to check how many weights in a 70b F32 model would be squashed when converted to F16 (spoiler, it's shockingly few)
The reason for this comparison is that it should represent the same percentage of squishing as bf16 to fp16
Had claude make me a script, using the new Reflection-70B, and these are the results:
Total weights: 70553706496
Fully representable: 70530215524
Squashed: 23490972
Percentage squashed: 0.03%
0.03%!!!!
A couple things to note, this uses a roundtrip of F32 -> F16 -> F32 and then torch.isclose to account for rounding errors that come up by the very nature of extremely accurate numbers, but it uses VERY small tolerances (rtol=1e-5, atol=1e-8)
This is also examining EVERY weight that was stored at F32, and for most layers I was somewhere between 0% and 0.03% of weights being squashed, no major outliers.
Overall, I feel even safer converting to F16 for llama.cpp, the extremely small number of weights that fall outside the range are likely so small that they don't actually play a role in the final output of the model at inference anyways.
The reason for this comparison is that it should represent the same percentage of squishing as bf16 to fp16
Had claude make me a script, using the new Reflection-70B, and these are the results:
Total weights: 70553706496
Fully representable: 70530215524
Squashed: 23490972
Percentage squashed: 0.03%
0.03%!!!!
A couple things to note, this uses a roundtrip of F32 -> F16 -> F32 and then torch.isclose to account for rounding errors that come up by the very nature of extremely accurate numbers, but it uses VERY small tolerances (rtol=1e-5, atol=1e-8)
This is also examining EVERY weight that was stored at F32, and for most layers I was somewhere between 0% and 0.03% of weights being squashed, no major outliers.
Overall, I feel even safer converting to F16 for llama.cpp, the extremely small number of weights that fall outside the range are likely so small that they don't actually play a role in the final output of the model at inference anyways.
-
- 20 replies
Post
4792
@victor
(is this the only way to "DM" on HF?)
Had a funny thought, would it be at all possible to rework what shows up on our personal HF page?
Picture this: I upload a model to an organization, someone who follows me now has no idea that I've uploaded a model or to where, unless they also watch those repos (which also floods them with other notifications)
What if our main Huggingface page was a collection of both models that we've uploaded specifically to our profile, as well as models we've uploaded to organizations? That way it would all be contained in one central followable location, and I wouldn't have concerns about losing followership if I wanted to upload to an organization all of a sudden.
Had a funny thought, would it be at all possible to rework what shows up on our personal HF page?
Picture this: I upload a model to an organization, someone who follows me now has no idea that I've uploaded a model or to where, unless they also watch those repos (which also floods them with other notifications)
What if our main Huggingface page was a collection of both models that we've uploaded specifically to our profile, as well as models we've uploaded to organizations? That way it would all be contained in one central followable location, and I wouldn't have concerns about losing followership if I wanted to upload to an organization all of a sudden.
-
-
- 3 replies
Post
10101
So turns out I've been spreading a bit of misinformation when it comes to imatrix in llama.cpp
It starts true; imatrix runs the model against a corpus of text and tracks the activation of weights to determine which are most important
However what the quantization then does with that information is where I was wrong.
I think I made the accidental connection between imatrix and exllamav2's measuring, where ExLlamaV2 decides how many bits to assign to which weight depending on the goal BPW
Instead, what llama.cpp with imatrix does is it attempts to select a scale for a quantization block that most accurately returns the important weights to their original values, ie minimizing the dequantization error based on the importance of activations
The mildly surprising part is that it actually just does a relatively brute force search, it picks a bunch of scales and tries each and sees which one results in the minimum error for weights deemed important in the group
But yeah, turns out, the quantization scheme is always the same, it's just that the scaling has a bit more logic to it when you use imatrix
Huge shoutout to @compilade for helping me wrap my head around it - feel free to add/correct as well if I've messed something up
It starts true; imatrix runs the model against a corpus of text and tracks the activation of weights to determine which are most important
However what the quantization then does with that information is where I was wrong.
I think I made the accidental connection between imatrix and exllamav2's measuring, where ExLlamaV2 decides how many bits to assign to which weight depending on the goal BPW
Instead, what llama.cpp with imatrix does is it attempts to select a scale for a quantization block that most accurately returns the important weights to their original values, ie minimizing the dequantization error based on the importance of activations
The mildly surprising part is that it actually just does a relatively brute force search, it picks a bunch of scales and tries each and sees which one results in the minimum error for weights deemed important in the group
But yeah, turns out, the quantization scheme is always the same, it's just that the scaling has a bit more logic to it when you use imatrix
Huge shoutout to @compilade for helping me wrap my head around it - feel free to add/correct as well if I've messed something up
- 5 replies
Post
6366
As some of you know, I try to convert models to either fp32 or bf16 depending on theirs size before doing imatrix and quantization
Today I decided to see if that matters, and the results have me.. for lack of a better word, perplexed
My setup:
Mistral Nemo Instruct 2407
- convert to FP32, calculate imatrix, quantize to Q8_0 and Q4_K_M
- convert to FP16, calculate imatrix, quantize to Q8_0 and Q4_K_M
I calculated the kld base from the FP32 model:
then calculated the divergence itself for each like so:
Q4_K_M from fp16 and fp32 were similar, trading blows across statistics, odd since i expected fp32 to be strictly better but it's not
Q8_0 is where things get weird. Despite each file being slightly different size, and the sha256sum of course being different, they each get *completely identical* scores, down to 6 decimal places of precision on the statistics.
How is this possible? Is there something I don't understand about llama.cpp that makes it always convert to fp16 before it does quantization? Am I wasting time using FP32/BF16??
Today I decided to see if that matters, and the results have me.. for lack of a better word, perplexed
My setup:
Mistral Nemo Instruct 2407
- convert to FP32, calculate imatrix, quantize to Q8_0 and Q4_K_M
- convert to FP16, calculate imatrix, quantize to Q8_0 and Q4_K_M
I calculated the kld base from the FP32 model:
./llama-perplexity -m /models/Mistral-Nemo-Instruct-2407-f32.gguf -f /training_data/wikitext-2-raw/wiki.test.raw --kl-divergence-base /training_data/mistral-nemo-f32.kld -ngl 35 -fa -sm rowthen calculated the divergence itself for each like so:
./llama-perplexity -m /models/Mistral-Nemo-Instruct-2407-Q8_0.gguf -f /training_data/wikitext-2-raw/wiki.test.raw --kl-divergence-base /training_data/mistral-nemo-f32.kld --kl-divergence -ngl 50 -fa -sm rowQ4_K_M from fp16 and fp32 were similar, trading blows across statistics, odd since i expected fp32 to be strictly better but it's not
Q8_0 is where things get weird. Despite each file being slightly different size, and the sha256sum of course being different, they each get *completely identical* scores, down to 6 decimal places of precision on the statistics.
How is this possible? Is there something I don't understand about llama.cpp that makes it always convert to fp16 before it does quantization? Am I wasting time using FP32/BF16??
-
- 17 replies
Lin-Chen
authored
2
papers
11 months ago
Post
6323
🚨Exciting news for the Multilingual Synthetic Data Community!🚨
I’ve taken inspiration from the MAGPIE paper on Llama-3-8B-instruct and extended its capabilities. Here’s what’s new!
🗞 The MAGPIE paper showcased that if you use the instruction-tuned version (
🤔 While reading a script by Sebastian Raschka, PhD, I wondered: Could these advancements be replicated in other languages? Specifically, could they benefit non-English datasets?
🎉 And the answer is YES! At least for Spanish. I've successfully adapted the techniques for Spanish, proving the model's flexibility and multilingual capabilities.
👩💻 To make this accessible, I created a basic script (heavily inspired by the Sebastian Raschka one) that allows you to generate similar datasets using
[Script](https://gist.github.com/mrm8488/4650a5e3cc45523798a527a3446eb312)
🔍 Explore the datasets 📚 generated using our new script!
- [Llama-3-8B](https://huggingface.co/datasets/mrm8488/dataset_llama3_5000_samples_es_4231_filtered)
- [Phi-3-medium](https://huggingface.co/datasets/mrm8488/dataset_phi3-medium_5000_samples_es_3906_filtered)
- [Phi-3-mini](https://huggingface.co/datasets/mrm8488/dataset_phi3_5000_samples_es_3282_filtered)
Note: These datasets have basic filtering. Apply additional quality filters before using them to fine-tune large language models.
Inspiration and base script:
https://github.com/rasbt/LLMs-from-scratch/blob/main/ch07/05_dataset-generation/llama3-ollama.ipynb
https://www.linkedin.com/feed/update/urn:li:activity:7210982019751661568/
I’ve taken inspiration from the MAGPIE paper on Llama-3-8B-instruct and extended its capabilities. Here’s what’s new!
🗞 The MAGPIE paper showcased that if you use the instruction-tuned version (
Llama-3-8B-instruct) to generate synthetic instructions and then fine-tune the base version (Llama-3-8B) on this dataset, you can improve even the it-tuned version🤔 While reading a script by Sebastian Raschka, PhD, I wondered: Could these advancements be replicated in other languages? Specifically, could they benefit non-English datasets?
🎉 And the answer is YES! At least for Spanish. I've successfully adapted the techniques for Spanish, proving the model's flexibility and multilingual capabilities.
👩💻 To make this accessible, I created a basic script (heavily inspired by the Sebastian Raschka one) that allows you to generate similar datasets using
ollama models (initially phi and llama3) automatically and upload it to the Hugging Face Hub
🔍 Explore the datasets 📚 generated using our new script!
- [Llama-3-8B](https://huggingface.co/datasets/mrm8488/dataset_llama3_5000_samples_es_4231_filtered)
- [Phi-3-medium](https://huggingface.co/datasets/mrm8488/dataset_phi3-medium_5000_samples_es_3906_filtered)
- [Phi-3-mini](https://huggingface.co/datasets/mrm8488/dataset_phi3_5000_samples_es_3282_filtered)
Note: These datasets have basic filtering. Apply additional quality filters before using them to fine-tune large language models.
Inspiration and base script:
https://github.com/rasbt/LLMs-from-scratch/blob/main/ch07/05_dataset-generation/llama3-ollama.ipynb
https://www.linkedin.com/feed/update/urn:li:activity:7210982019751661568/
- 7 replies