maya-research/maya1
Text-to-Speech
•
3B
•
Updated
•
60.1k
•
•
794
Felix Fischer
FlipTip
AI & ML interests
None yet
Recent Activity
liked
a model
25 days ago
maya-research/maya1
upvoted
an
article
4 months ago
TextQuests: How Good are LLMs at Text-Based Video Games?
Organizations
None yet
liked
a
model
25 days ago
reacted to
codelion's
post with 🔥
3 months ago
Post
5276
I wanted to share a technique that's been working really well for recovering performance after INT4 quantization.
Typically, quantizing the LLM to INT4 (unlike say INT8) for inference can incur some accuracy loss. Instead of accepting the quality loss, we used the FP16 model as a teacher to train a tiny LoRA adapter (rank=16) for the quantized model. The cool part: the model generates its own training data using the Magpie technique so no external datasets needed. This is critical because we want to remain as much as possible in the distribution of the model's natural responses.
Last year Apple's foundational models paper (https://arxiv.org/pdf/2407.21075) had proposed a similar technique and found "By using accuracy-recovery LoRA adapters with only rank 16, Alpaca win rate can be improved by 7-18%, GMS8K accuracy is boosted by 5-10%." (page 47).
We saw similar results on Qwen3-0.6B:
Perplexity: 2.40 → 2.09 (only 5.7% degradation from FP16 baseline)
Memory: Only 0.28GB vs 1.0GB for FP16 (75% reduction)
Speed: 3.0x faster inference than FP16
Quality: Generates correct, optimized code solutions
- Pre-trained adapter: codelion/Qwen3-0.6B-accuracy-recovery-lora
- GitHub repo: https://github.com/codelion/ellora
Happy to answer questions about the implementation or help anyone trying to replicate this. The key insight is that quantization errors are systematic and learnable - a small adapter can bridge the gap without negating the benefits of quantization.
Has anyone else experimented with self-distillation for quantization recovery? Would love to hear about different approaches!
Typically, quantizing the LLM to INT4 (unlike say INT8) for inference can incur some accuracy loss. Instead of accepting the quality loss, we used the FP16 model as a teacher to train a tiny LoRA adapter (rank=16) for the quantized model. The cool part: the model generates its own training data using the Magpie technique so no external datasets needed. This is critical because we want to remain as much as possible in the distribution of the model's natural responses.
Last year Apple's foundational models paper (https://arxiv.org/pdf/2407.21075) had proposed a similar technique and found "By using accuracy-recovery LoRA adapters with only rank 16, Alpaca win rate can be improved by 7-18%, GMS8K accuracy is boosted by 5-10%." (page 47).
We saw similar results on Qwen3-0.6B:
Perplexity: 2.40 → 2.09 (only 5.7% degradation from FP16 baseline)
Memory: Only 0.28GB vs 1.0GB for FP16 (75% reduction)
Speed: 3.0x faster inference than FP16
Quality: Generates correct, optimized code solutions
- Pre-trained adapter: codelion/Qwen3-0.6B-accuracy-recovery-lora
- GitHub repo: https://github.com/codelion/ellora
Happy to answer questions about the implementation or help anyone trying to replicate this. The key insight is that quantization errors are systematic and learnable - a small adapter can bridge the gap without negating the benefits of quantization.
Has anyone else experimented with self-distillation for quantization recovery? Would love to hear about different approaches!
upvoted
an
article
4 months ago
Article
TextQuests: How Good are LLMs at Text-Based Video Games?
•
36
liked
a
model
4 months ago
Suggestion: publishing (parts of the) training data
2
#5 opened 7 months ago
by
FlipTip
Thank you for your work
reacted to
wolfram's
post with 🤗
7 months ago
Post
7734
Finally finished my extensive **Qwen 3 evaluations** across a range of formats and quantisations, focusing on **MMLU-Pro** (Computer Science).
A few take-aways stood out - especially for those interested in local deployment and performance trade-offs:
1️⃣ **Qwen3-235B-A22B** (via Fireworks API) tops the table at **83.66%** with ~55 tok/s.
2️⃣ But the **30B-A3B Unsloth** quant delivered **82.20%** while running locally at ~45 tok/s and with zero API spend.
3️⃣ The same Unsloth build is ~5x faster than Qwen's **Qwen3-32B**, which scores **82.20%** as well yet crawls at <10 tok/s.
4️⃣ On Apple silicon, the **30B MLX** port hits **79.51%** while sustaining ~64 tok/s - arguably today's best speed/quality trade-off for Mac setups.
5️⃣ The **0.6B** micro-model races above 180 tok/s but tops out at **37.56%** - that's why it's not even on the graph (50 % performance cut-off).
All local runs were done with LM Studio on an M4 MacBook Pro, using Qwen's official recommended settings.
**Conclusion:** Quantised 30B models now get you ~98 % of frontier-class accuracy - at a fraction of the latency, cost, and energy. For most local RAG or agent workloads, they're not just good enough - they're the new default.
Well done, Qwen - you really whipped the llama's ass! And to OpenAI: for your upcoming open model, please make it MoE, with toggleable reasoning, and release it in many sizes. *This* is the future!
A few take-aways stood out - especially for those interested in local deployment and performance trade-offs:
1️⃣ **Qwen3-235B-A22B** (via Fireworks API) tops the table at **83.66%** with ~55 tok/s.
2️⃣ But the **30B-A3B Unsloth** quant delivered **82.20%** while running locally at ~45 tok/s and with zero API spend.
3️⃣ The same Unsloth build is ~5x faster than Qwen's **Qwen3-32B**, which scores **82.20%** as well yet crawls at <10 tok/s.
4️⃣ On Apple silicon, the **30B MLX** port hits **79.51%** while sustaining ~64 tok/s - arguably today's best speed/quality trade-off for Mac setups.
5️⃣ The **0.6B** micro-model races above 180 tok/s but tops out at **37.56%** - that's why it's not even on the graph (50 % performance cut-off).
All local runs were done with LM Studio on an M4 MacBook Pro, using Qwen's official recommended settings.
**Conclusion:** Quantised 30B models now get you ~98 % of frontier-class accuracy - at a fraction of the latency, cost, and energy. For most local RAG or agent workloads, they're not just good enough - they're the new default.
Well done, Qwen - you really whipped the llama's ass! And to OpenAI: for your upcoming open model, please make it MoE, with toggleable reasoning, and release it in many sizes. *This* is the future!
-
- 4 replies
Problem with web search
👍
2
3
#716 opened 7 months ago
by
DaniFera
liked
a
Space
8 months ago
reacted to
onekq's
post with 🔥
8 months ago
Post
1560
This is bitter lesson 2.0
https://storage.googleapis.com/deepmind-media/Era-of-Experience%20/The%20Era%20of%20Experience%20Paper.pdf
If this reads too lofty to you, consider some low-hanging fruits. Experiences here are reward signals we send to LLMs, e.g. human score in RLHF, verification in AlphaProof, or test results for code generation.
RFT (reinforced finetuning) will become main stream, and IMO make LLMs behave more like agents.
https://storage.googleapis.com/deepmind-media/Era-of-Experience%20/The%20Era%20of%20Experience%20Paper.pdf
If this reads too lofty to you, consider some low-hanging fruits. Experiences here are reward signals we send to LLMs, e.g. human score in RLHF, verification in AlphaProof, or test results for code generation.
RFT (reinforced finetuning) will become main stream, and IMO make LLMs behave more like agents.
- 2 replies
reacted to
giux78's
post with 👀
8 months ago
Post
2392
LLAMA4 release highlight the importance of political and social bias. According to their own evaluation described in the release blog post:
- Refusals on contentious prompts dropped from 7% (hashtag#LLAMA 3.3) to under 2%
- Unequal response refusals are now under 1%
- Political lean bias is said to be halved compared to hashtag#LLaMA 3.3 and comparable to Grok
However, we @efederici @mferraretto @FinancialSupport and I released some weeks ago an independent open source benchmark called Propaganda to measure political bias in LLMs: https://github.com/mii-llm/propaganda
In the chart below, we evaluated multiple leading models on the basis of ratings across a range of prompts designed to expose ideological leanings.
Despite Meta’s stated neutrality goals, LLAMA4 ranks at the very top in terms of total ratings aligned with a clear ideological bias. The models were tested on their ability to respond even-handedly to politically sensitive prompts. LLaMA 4 scored even higher than models known for strong alignment policies like GPT-4o.
LLMs may be refusing less, but they still show bias through content framing. This suggests that refusal rates alone are not a sufficient measure of ideological bias. Relying solely on internal evaluations from AI labs also raises concerns about transparency and objectivity.
- Refusals on contentious prompts dropped from 7% (hashtag#LLAMA 3.3) to under 2%
- Unequal response refusals are now under 1%
- Political lean bias is said to be halved compared to hashtag#LLaMA 3.3 and comparable to Grok
However, we @efederici @mferraretto @FinancialSupport and I released some weeks ago an independent open source benchmark called Propaganda to measure political bias in LLMs: https://github.com/mii-llm/propaganda
In the chart below, we evaluated multiple leading models on the basis of ratings across a range of prompts designed to expose ideological leanings.
Despite Meta’s stated neutrality goals, LLAMA4 ranks at the very top in terms of total ratings aligned with a clear ideological bias. The models were tested on their ability to respond even-handedly to politically sensitive prompts. LLaMA 4 scored even higher than models known for strong alignment policies like GPT-4o.
LLMs may be refusing less, but they still show bias through content framing. This suggests that refusal rates alone are not a sufficient measure of ideological bias. Relying solely on internal evaluations from AI labs also raises concerns about transparency and objectivity.
reacted to
m-ric's
post with 🔥
9 months ago
Post
10017
Introducing 𝗼𝗽𝗲𝗻 𝗗𝗲𝗲𝗽-𝗥𝗲𝘀𝗲𝗮𝗿𝗰𝗵 by Hugging Face! 💥
OpenAI's latest agentic app Deep Research seems really good... But it's closed, as usual.
⏱️ So with a team of cracked colleagues, we set ourselves a 24hours deadline to replicate and open-source Deep Research! ⏱️
➡️ We built open-Deep-Research, an entirely open agent that can: navigate the web autonomously, scroll and search through pages, download and manipulate files, run calculation on data...
We aimed for the best performance: are the agent's answers really rigorous?
On GAIA benchmark, Deep Research had 67% accuracy on the validation set.
➡️ open Deep Research is at 55% (powered by o1), it is:
- the best pass@1 solution submitted
- the best open solution 💪💪
And it's only getting started ! Please jump in, drop PRs, and let's bring it to the top !
Read the blog post 👉 https://huggingface.co/blog/open-deep-research
OpenAI's latest agentic app Deep Research seems really good... But it's closed, as usual.
⏱️ So with a team of cracked colleagues, we set ourselves a 24hours deadline to replicate and open-source Deep Research! ⏱️
➡️ We built open-Deep-Research, an entirely open agent that can: navigate the web autonomously, scroll and search through pages, download and manipulate files, run calculation on data...
We aimed for the best performance: are the agent's answers really rigorous?
On GAIA benchmark, Deep Research had 67% accuracy on the validation set.
➡️ open Deep Research is at 55% (powered by o1), it is:
- the best pass@1 solution submitted
- the best open solution 💪💪
And it's only getting started ! Please jump in, drop PRs, and let's bring it to the top !
Read the blog post 👉 https://huggingface.co/blog/open-deep-research
reacted to
m-ric's
post with 🔥👍
9 months ago
Post
3150
Less is More for Reasoning (LIMO): a 32B model fine-tuned with 817 examples can beat o1-preview on math reasoning! 🤯
Do we really need o1's huge RL procedure to see reasoning emerge? It seems not.
Researchers from Shanghai Jiaotong University just demonstrated that carefully selected examples can boost math performance in large language models using SFT —no huge datasets or RL procedures needed.
Their procedure allows Qwen2.5-32B-Instruct to jump from 6.5% to 57% on AIME and from 59% to 95% on MATH, while using only 1% of the data in previous approaches.
⚡ The Less-is-More Reasoning Hypothesis:
‣ Minimal but precise examples that showcase optimal reasoning patterns matter more than sheer quantity
‣ Pre-training knowledge plus sufficient computational resources at inference levels up math skills
➡️ Core techniques:
‣ High-quality reasoning chains with self-verification steps
‣ 817 handpicked problems that encourage deeper reasoning
‣ Enough inference-time computation to allow extended reasoning
💪 Efficiency gains:
‣ Only 817 examples instead of 100k+
‣ 40.5% absolute improvement across 10 diverse benchmarks, outperforming models trained on 100x more data
This really challenges the notion that SFT leads to memorization rather than generalization! And opens up reasoning to GPU-poor researchers 🚀
Read the full paper here 👉 LIMO: Less is More for Reasoning (2502.03387)
Do we really need o1's huge RL procedure to see reasoning emerge? It seems not.
Researchers from Shanghai Jiaotong University just demonstrated that carefully selected examples can boost math performance in large language models using SFT —no huge datasets or RL procedures needed.
Their procedure allows Qwen2.5-32B-Instruct to jump from 6.5% to 57% on AIME and from 59% to 95% on MATH, while using only 1% of the data in previous approaches.
⚡ The Less-is-More Reasoning Hypothesis:
‣ Minimal but precise examples that showcase optimal reasoning patterns matter more than sheer quantity
‣ Pre-training knowledge plus sufficient computational resources at inference levels up math skills
➡️ Core techniques:
‣ High-quality reasoning chains with self-verification steps
‣ 817 handpicked problems that encourage deeper reasoning
‣ Enough inference-time computation to allow extended reasoning
💪 Efficiency gains:
‣ Only 817 examples instead of 100k+
‣ 40.5% absolute improvement across 10 diverse benchmarks, outperforming models trained on 100x more data
This really challenges the notion that SFT leads to memorization rather than generalization! And opens up reasoning to GPU-poor researchers 🚀
Read the full paper here 👉 LIMO: Less is More for Reasoning (2502.03387)
liked
a
model
10 months ago
updated
a
collection
about 1 year ago
liked
a
model
about 1 year ago
