kshitijthakkar/repro-grpo-success-amplification-artifacts
68.2 kB
Kshitij Thakkar PRO
kshitijthakkar
AI & ML interests
Building the evaluation and observability layer for AI.
Creator of TraceVerse—turning real-world LLM interactions into datasets, benchmarks, and cost-efficient model insights.
Recent Activity
updated a collection 6 minutes ago
ICML 2026 Reproductions - agent-repro challenge updated a Space 7 minutes ago
kshitijthakkar/repro-grpo-success-amplification published a Space 7 minutes ago
kshitijthakkar/repro-grpo-success-amplificationOrganizations
buckets 20
kshitijthakkar/repro-disentanglement-identifiability-artifacts
26.4 kB
kshitijthakkar/repro-treatment-allocation-risk-artifacts
86.8 kB
kshitijthakkar/repro-mfvi-predictive-variance-artifacts
113 kB
kshitijthakkar/repro-docalculus-derivation-graphs-artifacts
104 kB
kshitijthakkar/repro-deaction-misactbench-artifacts
1.75 MB
Articles 6
Article
The Mind of Tashi: making a 200M-active model's reasoning *the game*
Article
2
Scaling Mixture of Experts: Architecture Search for Billion-Parameter Language Models
Trackio logbooks: claim-by-claim reproductions of ICML 2026 papers for the Agent Reproducibility Challenge.
-
Repro - MemEvolve: Meta-Evolution of Agent Memory Systems
🧬Collaborate with an AI agent to manage a shared experiment logbook
-
Repro - TruthRL: Incentivizing Truthful LLMs via Reinforcement Learning
🎯Collaborate with an AI agent using a shared logbook
-
Repro - How to Correctly Report LLM-as-a-Judge Evaluations
🎯Log and share LLM evaluation findings in a collaborative notebook
-
Repro - Dependence-Aware Label Aggregation via Ising Models
🧲Collaborate with an AI agent via a shared logbook
Qwen3.6-35B-A3B carved by expert importance to fit one consumer GPU. No training, no calibration. 489 tok/s on a laptop 5090.
-
kshitijthakkar/Kirigami-Qwen3.6-20B-A3B-NVFP4
Text Generation • 14B • Updated • 35 -
kshitijthakkar/Kirigami-Qwen3.6-24B-A3B-NVFP4
Text Generation • 16B • Updated • 18 -
kshitijthakkar/Kirigami-Qwen3.6-28B-A3B-NVFP4
Text Generation • 19B • Updated • 12 -
Kirigami Journey
🪷How we carved a 35B MoE to fit a 24GB GPU — zero training
Trackio logbooks: claim-by-claim reproductions of ICML 2026 papers for the Agent Reproducibility Challenge.
-
Repro - MemEvolve: Meta-Evolution of Agent Memory Systems
🧬Collaborate with an AI agent to manage a shared experiment logbook
-
Repro - TruthRL: Incentivizing Truthful LLMs via Reinforcement Learning
🎯Collaborate with an AI agent using a shared logbook
-
Repro - How to Correctly Report LLM-as-a-Judge Evaluations
🎯Log and share LLM evaluation findings in a collaborative notebook
-
Repro - Dependence-Aware Label Aggregation via Ising Models
🧲Collaborate with an AI agent via a shared logbook
Qwen3.6-35B-A3B carved by expert importance to fit one consumer GPU. No training, no calibration. 489 tok/s on a laptop 5090.
-
kshitijthakkar/Kirigami-Qwen3.6-20B-A3B-NVFP4
Text Generation • 14B • Updated • 35 -
kshitijthakkar/Kirigami-Qwen3.6-24B-A3B-NVFP4
Text Generation • 16B • Updated • 18 -
kshitijthakkar/Kirigami-Qwen3.6-28B-A3B-NVFP4
Text Generation • 19B • Updated • 12 -
Kirigami Journey
🪷How we carved a 35B MoE to fit a 24GB GPU — zero training
spaces 34
pinned
Running
Agents
GuardianTails
🐾
Pet Health Intelligence Platform
Running
Reproducing GRPO's Loss, Dynamics, and Success Amplification (arXiv:2503.06639)
📈
Running
Repro - Unsupervised Disentanglement Without Compromises
🔀
Explore experiment logs and sync with a coding agent
Running
Repro - Learning Treatment Allocations with Risk Control Under Partial Identifiability
⚖
Collaborate on a research logbook with an AI agent
Running
Repro - Gaussian Mean Field Variational Inference can Overestimate Predictive Variance
📉
Browse research logbooks and collaborate with an AI agent
Running
Repro - Do-Calculus Derivation Graphs
🔗
Explore and share do‑calculus derivation graphs
models 147
kshitijthakkar/Kirigami-Qwen3.6-28B-A3B-NVFP4
Text Generation • 19B • Updated • 12
kshitijthakkar/Kirigami-Qwen3.6-20B-A3B-NVFP4
Text Generation • 14B • Updated • 35
kshitijthakkar/Kirigami-Qwen3.6-24B-A3B-NVFP4
Text Generation • 16B • Updated • 18
kshitijthakkar/deepseek-v4-mini-300M-sft-v2
Text Generation • 0.3B • Updated • 289
kshitijthakkar/deepseek-v4-mini-300M-sft-v1
Text Generation • 0.3B • Updated • 281
kshitijthakkar/deepseek-v4-mini-300M-cpt-fineweb
Text Generation • 0.3B • Updated • 282
kshitijthakkar/deepseek-v4-mini-300M-recovered
Text Generation • 0.3B • Updated • 31 • 1
kshitijthakkar/deepseek-v4-mini-300M-recovered-h100
Text Generation • 0.3B • Updated • 3
kshitijthakkar/deepseek-v4-mini-300M-recovered-wip
Text Generation • 0.3B • Updated • 6
kshitijthakkar/deepseek-v4-mini-300M-from-flash-sft-test-lora
Updated • 1
datasets 419
kshitijthakkar/EEE_datastore
Viewer • Updated • 4.89k • 16
kshitijthakkar/Nemotron-Personas-India
Viewer • Updated • 3M • 182
kshitijthakkar/smoltrace-leaderboard
Viewer • Updated • 108 • 1.1k
kshitijthakkar/smoltrace-metrics-20260424_122614
Viewer • Updated • 1 • 11
kshitijthakkar/smoltrace-traces-20260424_122614
Viewer • Updated • 2 • 8
kshitijthakkar/smoltrace-results-20260424_122614
Viewer • Updated • 2 • 14
kshitijthakkar/smoltrace-metrics-20260424_112312
Viewer • Updated • 1 • 18
kshitijthakkar/smoltrace-traces-20260424_112312
Viewer • Updated • 2 • 17
kshitijthakkar/smoltrace-results-20260424_112312
Viewer • Updated • 2 • 10
kshitijthakkar/smoltrace-metrics-20260424_111528
Viewer • Updated • 1 • 17