Ask Timestral · Bologna bikes API
🔮
Hourly bike forecasts for Bologna, next 7 days (API)
Massimiliano PRO
Bayway
AI & ML interests
None yet
Recent Activity
updated a Space 1 day ago
Bayway/due-torri-due-ruote-api updated a Space 1 day ago
Bayway/due-torri-due-ruote reacted to gianlucamazza's post with 🚀 2 days ago
A live demo of emotional-memory 0.18.1: chat, watch the affective state move, and type "recall X" to see memories ranked by meaning plus mood. It shows how the mechanism works, not that it wins in general. The Scope and limits section lists the benchmarks where plain cosine beats it.
https://huggingface.co/spaces/gianlucamazza/emotional-memory-demoOrganizations
None yet
updated 2
Spaces 1 day ago
reacted to gianlucamazza's post with 🚀 2 days ago
Post
89
A live demo of emotional-memory 0.18.1: chat, watch the affective state move, and type "recall X" to see memories ranked by meaning plus mood. It shows how the mechanism works, not that it wins in general. The Scope and limits section lists the benchmarks where plain cosine beats it.
gianlucamazza/emotional-memory-demo
gianlucamazza/emotional-memory-demo
- 13 replies
liked a
Space 2 days ago
posted an update 2 days ago
Post
1,940
Italian Open Time Series: 46 public Italian series for zero-shot forecasting, and a first live test of Timestral
Italian data is almost absent from the corpora that time-series foundation models are trained and benchmarked on. So I collected 46 open series (CC0/CC BY): smog in Milan, bikes in Bologna, tourism, energy, earthquakes, deaths, Italian Wikipedia. Gaps stay NaN, and a tasks.yaml makes every window reproducible with fev.
Seven models, zero-shot, 19 tasks (Timestral 5M to 209M, Chronos-2, TimesFM 3.0, Toto 2.0). Skill vs seasonal naive, fev-bench style:
• Backtest, windows ending 14 Oct, Timestral's release day: Toto 2.0 28.4, Timestral-209M 28.2
• Live, data published after that day (6 tasks, ~1 week): Timestral-209M 32.2, Timestral-123M 30.9, TimesFM 3.0 30.3, Toto 2.0 27.7
• Monthly consumer prices of electricity and gas: plain AutoETS beats every foundation model
The live round is small; I'll re-run it every week.
Two surprises.
Weather beats the calendar. On a Bologna bike counter, giving Timestral the observed weather of the forecast days (a perfect forecast) cut the error by 17%. Holidays and bridge days alone: 4%.
Daylight saving time. The model works in UTC; cyclists don't. Across the 25 Oct clock change my first version put Monday's rush hour at 5 am, until I moved the series to Italian wall-clock time.
For Bologna there's a small page: pick a street and a day in the next week for an hourly bike forecast, computed live from the city's counts and Open-Meteo, plus a map of counters and air-quality stations.
Dataset: Bayway/italian-open-timeseries
Bologna: Bayway/due-torri-due-ruote
Not included: GME wholesale power prices and Terna load data. Their terms forbid redistribution, and GME bans automated extraction for any purpose, AI training included.
Italian data is almost absent from the corpora that time-series foundation models are trained and benchmarked on. So I collected 46 open series (CC0/CC BY): smog in Milan, bikes in Bologna, tourism, energy, earthquakes, deaths, Italian Wikipedia. Gaps stay NaN, and a tasks.yaml makes every window reproducible with fev.
Seven models, zero-shot, 19 tasks (Timestral 5M to 209M, Chronos-2, TimesFM 3.0, Toto 2.0). Skill vs seasonal naive, fev-bench style:
• Backtest, windows ending 14 Oct, Timestral's release day: Toto 2.0 28.4, Timestral-209M 28.2
• Live, data published after that day (6 tasks, ~1 week): Timestral-209M 32.2, Timestral-123M 30.9, TimesFM 3.0 30.3, Toto 2.0 27.7
• Monthly consumer prices of electricity and gas: plain AutoETS beats every foundation model
The live round is small; I'll re-run it every week.
Two surprises.
Weather beats the calendar. On a Bologna bike counter, giving Timestral the observed weather of the forecast days (a perfect forecast) cut the error by 17%. Holidays and bridge days alone: 4%.
Daylight saving time. The model works in UTC; cyclists don't. Across the 25 Oct clock change my first version put Monday's rush hour at 5 am, until I moved the series to Italian wall-clock time.
For Bologna there's a small page: pick a street and a day in the next week for an hourly bike forecast, computed live from the city's counts and Open-Meteo, plus a map of counters and air-quality stations.
Dataset: Bayway/italian-open-timeseries
Bologna: Bayway/due-torri-due-ruote
Not included: GME wholesale power prices and Terna load data. Their terms forbid redistribution, and GME bans automated extraction for any purpose, AI training included.
published 2
Spaces 2 days ago
updated a
dataset 2 days ago
published a
dataset 2 days ago
reacted to jiteshkprasad's post with 🔥 4 days ago
Post
2,914
Last week we released Gargi, a 384M English language model trained from scratch on about 17B tokens, roughly 1,000× less training data than Qwen2.5-0.5B (~18T tokens, as published by its authors).
What we measured (same harness for every model: lm-evaluation-harness, 0-shot, greedy, each model's start token):
- Base model vs Qwen2.5-0.5B: ARC-Easy 62.8 vs 64.7, ARC-Challenge 32.3 vs 32.4, WinoGrande 54.6 vs 56.1.
- Chat model: thinking on or off; GSM8K 36.0 with thinking on.
- Long inputs: needle-in-a-haystack recall 100% at 8K, 93% at 16K and 90% at 20K tokens, so use it up to about 16K.
Where it falls short: knowledge and everyday common sense still scale with data. MMLU is 25.5 vs 47.3, and it trails on HellaSwag and PIQA. It is not a maths specialist, and it is not safety-tuned: put a moderation layer in front of any public use.
It runs on a laptop: GGUF (Q8_0 at 410 MB, Q4_K_M, f16) and MLX builds.
Apache-2.0 ·
weights: udgamlabs/Gargi
try it: https://udgamlabs.com/models/gargi
code: https://github.com/udgamlabs/gargi
Named after Gargi Vachaknavi, the Vedic philosopher remembered for asking the hardest questions.
What we measured (same harness for every model: lm-evaluation-harness, 0-shot, greedy, each model's start token):
- Base model vs Qwen2.5-0.5B: ARC-Easy 62.8 vs 64.7, ARC-Challenge 32.3 vs 32.4, WinoGrande 54.6 vs 56.1.
- Chat model: thinking on or off; GSM8K 36.0 with thinking on.
- Long inputs: needle-in-a-haystack recall 100% at 8K, 93% at 16K and 90% at 20K tokens, so use it up to about 16K.
Where it falls short: knowledge and everyday common sense still scale with data. MMLU is 25.5 vs 47.3, and it trails on HellaSwag and PIQA. It is not a maths specialist, and it is not safety-tuned: put a moderation layer in front of any public use.
It runs on a laptop: GGUF (Q8_0 at 410 MB, Q4_K_M, f16) and MLX builds.
ollama run hf.co/udgamlabs/Gargi-GGUF:Q8_0Apache-2.0 ·
weights: udgamlabs/Gargi
try it: https://udgamlabs.com/models/gargi
code: https://github.com/udgamlabs/gargi
Named after Gargi Vachaknavi, the Vedic philosopher remembered for asking the hardest questions.
- 21 replies
reacted to gianlucamazza's post with 🔥 13 days ago
Post
46
I released my own benchmark sets behind emotional-memory, my library for mood-aware recall in LLM agents. The card opens with the limits: the main set is built to favour the method, and the negative results are listed up front, e.g. on LoCoMo AFT scores F1 0.168 against 0.271 for a naive RAG baseline. Use it to check where affect-conditioned retrieval helps and where it doesn't.
gianlucamazza/emotional-memory-benchmarks
gianlucamazza/emotional-memory-benchmarks
reacted to gianlucamazza's post with 🚀 13 days ago
Post
4,556
Upstream models my open-source projects run or my articles discuss. Each note says which it is. thenlper/gte-small is real use: RAGFS embeds file chunks with it locally. The two retrieval models from my RAG article, a cross-encoder re-ranker and a bi-encoder, are examples, labelled as such, not claims of production use.
gianlucamazza/models-used-in-my-projects-6aca0b3201f38f4c264ac9bf
gianlucamazza/models-used-in-my-projects-6aca0b3201f38f4c264ac9bf
updated a
model 16 days ago
posted an update 18 days ago
Post
84
JEV-27B-VL now runs on a 32 GB Mac 🍎
AutoTrust's JEV answers typed questions (yes/no, pick one of up to 256 options, a 0–5 score) about text and images with a calibrated probability per option, in a single forward pass. The model card asks for an 80 GB GPU; with MLX it peaks at 16.9 GB and takes ~1.5 s per decision on an M2 Pro.
What's in it:
• 4-bit Qwen3.8-27B base (byte-identical to mlx-community's) + JEV's System 1 LoRA kept unmerged in bf16 + a float32 decision head
• Same checkpoint = plain Qwen3.8-27B when the adapter is off (System 2)
• Images in the state, 2–256 options
Parity, measured against the official PyTorch math on JEV-9B: bf16 39/39 decisions, max probability diff 0.008 (inside the noise between the two official reference paths); 8-bit indistinguishable; 4-bit keeps every clear decision but moves probabilities up to 0.12. The 27B 4-bit build matches the outputs published on the JEV cards within 0.003. Full report in the repo.
Weights: Bayway/JEV-27B-VL-MLX-4bit
mlx-vlm support (under review): https://github.com/Blaizzy/mlx-vlm/pull/2463
All credit for the model to AutoTrust ( autotrust/JEV-27B-VL); thanks to the mlx-vlm and mlx-community folks for the decision API and the base quant.
AutoTrust's JEV answers typed questions (yes/no, pick one of up to 256 options, a 0–5 score) about text and images with a calibrated probability per option, in a single forward pass. The model card asks for an 80 GB GPU; with MLX it peaks at 16.9 GB and takes ~1.5 s per decision on an M2 Pro.
What's in it:
• 4-bit Qwen3.8-27B base (byte-identical to mlx-community's) + JEV's System 1 LoRA kept unmerged in bf16 + a float32 decision head
• Same checkpoint = plain Qwen3.8-27B when the adapter is off (System 2)
• Images in the state, 2–256 options
Parity, measured against the official PyTorch math on JEV-9B: bf16 39/39 decisions, max probability diff 0.008 (inside the noise between the two official reference paths); 8-bit indistinguishable; 4-bit keeps every clear decision but moves probabilities up to 0.12. The 27B 4-bit build matches the outputs published on the JEV cards within 0.003. Full report in the repo.
from mlx_vlm import load, predict
model, processor = load("Bayway/JEV-27B-VL-MLX-4bit")
predict(model, processor, "Customer: my card was charged twice for one coffee.",
{"team": {"type": "choice", "instructions": "Which team should handle this?",
"criteria": ["billing", "shipping", "tech support"]}})
# billing 0.998 · tech support 0.002Weights: Bayway/JEV-27B-VL-MLX-4bit
mlx-vlm support (under review): https://github.com/Blaizzy/mlx-vlm/pull/2463
All credit for the model to AutoTrust ( autotrust/JEV-27B-VL); thanks to the mlx-vlm and mlx-community folks for the decision API and the base quant.
published a
model 18 days ago
liked a
dataset 3 months ago
liked a
model 3 months ago
