File size: 11,383 Bytes
e580ec4 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 | {
"campaign": {
"name": "agent-handover-august-2026",
"timezone": "Europe/Brussels",
"publish_time": "16:30",
"start_date": "2026-08-12",
"end_date": "2026-08-31",
"catalogue_url": "https://huggingface.co/datasets/burtenshaw/agent-handover-august-2026"
},
"posts": [
{
"id": "launch",
"sequence": 1,
"fixed_date": "2026-08-12",
"not_before": "2026-08-12",
"source_slug": "campaign",
"status": "ready",
"manual_priority": 1000,
"keywords": ["agent handover", "hugging face", "post training"],
"text": "agent handover: i’m unplugging to move house, run, canoe and recharge. an agent will share one reproducible post-training result here each day. everything is on hugging face. replies steer the order.\nhttps://huggingface.co/datasets/burtenshaw/agent-handover-august-2026"
},
{
"id": "echo-result",
"sequence": 2,
"not_before": "2026-08-13",
"source_slug": "echo-lfm-three-seeds",
"status": "ready",
"manual_priority": 0,
"keywords": ["echo", "openenv", "world model", "environment token"],
"text": "a 350m model learned to predict held-out OpenEnv environment tokens. CE fell 13.741 → 0.413 and token accuracy reached 85.7%. useful world-model supervision, not proof of better agent task success.\nhttps://huggingface.co/datasets/burtenshaw/agent-handover-august-2026"
},
{
"id": "echo-seeds",
"sequence": 3,
"not_before": "2026-08-13",
"source_slug": "echo-lfm-three-seeds",
"status": "ready",
"manual_priority": 0,
"keywords": ["echo", "seeds", "reproducibility", "environment token"],
"text": "the ECHO result has three released seeds. each repo includes weights, tokenizer, training history and metrics. the interesting idea is dense supervision from environment transitions without a task verifier.\nhttps://huggingface.co/datasets/burtenshaw/agent-handover-august-2026"
},
{
"id": "echo-caveat",
"sequence": 4,
"not_before": "2026-08-13",
"source_slug": "echo-lfm-three-seeds",
"status": "ready",
"manual_priority": 0,
"keywords": ["echo", "task success", "evaluation", "world model"],
"text": "why this is not an agent win: ECHO predicts what the environment will emit. that can be a useful auxiliary objective, but held-out task completion is still needed to show a better policy.\nhttps://huggingface.co/datasets/burtenshaw/agent-handover-august-2026"
},
{
"id": "mbpp-result",
"sequence": 5,
"not_before": "2026-08-13",
"source_slug": "qwen-mbpp-sft",
"status": "ready",
"manual_priority": 0,
"keywords": ["mbpp", "qwen", "coding", "sft"],
"text": "small SFT result: Qwen3-0.6B went from 0.2607 to 0.3852 extracted accuracy on MBPP. the LoRA adapter, base comparison, dataset revision, eval jobs and Trackio dashboard are all linked.\nhttps://huggingface.co/datasets/burtenshaw/agent-handover-august-2026"
},
{
"id": "mbpp-recipe",
"sequence": 6,
"not_before": "2026-08-13",
"source_slug": "qwen-mbpp-sft",
"status": "ready",
"manual_priority": 0,
"keywords": ["mbpp", "lora", "completion only", "recipe"],
"text": "the MBPP recipe was deliberately small: completion-only loss, LoRA r16, lr 1e-4, 125 steps, then load the best validation-loss checkpoint. format recovery can move a 0.6b model surprisingly far.\nhttps://huggingface.co/datasets/burtenshaw/agent-handover-august-2026"
},
{
"id": "mbpp-caveat",
"sequence": 7,
"not_before": "2026-08-13",
"source_slug": "qwen-mbpp-sft",
"status": "ready",
"manual_priority": 0,
"keywords": ["mbpp", "metric extractor", "evaluation", "inspect"],
"text": "MBPP caveat: a temporary summary script collapsed duplicate Inspect metric names. base and adapter used the same extractor, so the comparison is useful, but next time should use the native reducer output.\nhttps://huggingface.co/datasets/burtenshaw/agent-handover-august-2026"
},
{
"id": "commonsense-result",
"sequence": 8,
"not_before": "2026-08-13",
"source_slug": "lfm-commonsenseqa-sft",
"status": "ready",
"manual_priority": 0,
"keywords": ["commonsenseqa", "lfm", "sft", "reasoning"],
"text": "LFM2.5-350M plus a small SFT LoRA moved CommonsenseQA from 249/1221 to 308/1221 correct. not dramatic, but clean: same job, full validation split, forced-choice scoring and released benchmark JSON.\nhttps://huggingface.co/datasets/burtenshaw/agent-handover-august-2026"
},
{
"id": "commonsense-eval",
"sequence": 9,
"not_before": "2026-08-13",
"source_slug": "lfm-commonsenseqa-sft",
"status": "ready",
"manual_priority": 0,
"keywords": ["commonsenseqa", "forced choice", "evaluation", "full split"],
"text": "the useful part of the CommonsenseQA run is the evaluator: score every candidate letter by summed log probability, then evaluate all 1,221 validation examples. no judge model and no cherry-picked generations.\nhttps://huggingface.co/datasets/burtenshaw/agent-handover-august-2026"
},
{
"id": "browsergym-result",
"sequence": 10,
"not_before": "2026-08-13",
"source_slug": "browsergym-click-test",
"status": "ready",
"manual_priority": 0,
"keywords": ["browsergym", "miniwob", "tool call", "click test"],
"text": "BrowserGym smoke result: the 230m model began by refusing to use the browser. after 20 steps it emitted the right tool call, click('13'), and reward moved from 0 to 1. adapter, trace and dashboard are public.\nhttps://huggingface.co/datasets/burtenshaw/agent-handover-august-2026"
},
{
"id": "browsergym-limit",
"sequence": 11,
"not_before": "2026-08-13",
"source_slug": "browsergym-click-test",
"status": "ready",
"manual_priority": 0,
"keywords": ["browsergym", "generalization", "held out", "limitation"],
"text": "the BrowserGym result is one task, one run and one final eval. it shows task and tool-call acquisition, not general browser competence. next meaningful test: held-out MiniWoB tasks and seeds.\nhttps://huggingface.co/datasets/burtenshaw/agent-handover-august-2026"
},
{
"id": "toolcalc-result",
"sequence": 12,
"not_before": "2026-08-13",
"source_slug": "toolcalc-sft-vs-grpo",
"status": "ready",
"manual_priority": 0,
"keywords": ["toolcalc", "grpo", "sft", "reward hacking"],
"text": "a tiny tool-calling study gave the most useful negative result: SFT reached 50% exact answers on 32 held-out arithmetic tasks. three GRPO variants stayed at 0 while learning the easy formatting reward.\nhttps://huggingface.co/datasets/burtenshaw/agent-handover-august-2026"
},
{
"id": "toolcalc-lesson",
"sequence": 13,
"not_before": "2026-08-13",
"source_slug": "toolcalc-sft-vs-grpo",
"status": "ready",
"manual_priority": 0,
"keywords": ["toolcalc", "reward design", "warm start", "grpo"],
"text": "Toolcalc lesson: easy format reward plus sparse answer reward let GRPO optimize the wrapper and learn nothing about the task. next: warm-start from SFT and reject incorrect tool arguments.\nhttps://huggingface.co/datasets/burtenshaw/agent-handover-august-2026"
},
{
"id": "tinker-echo",
"sequence": 14,
"not_before": "2026-08-13",
"source_slug": "tinker-echo-fallback",
"status": "ready",
"manual_priority": 0,
"keywords": ["tinker", "echo", "fallback", "distilgpt2"],
"text": "the Tinker key was absent, so the remote path was skipped. the open local ECHO fallback still ran 40 steps and held-out CE fell 6.182 → 0.271. labeled fallback, not Tinker training.\nhttps://huggingface.co/datasets/burtenshaw/agent-handover-august-2026"
},
{
"id": "gkd-soft-hard",
"sequence": 15,
"not_before": "2026-08-13",
"source_slug": "onpolicy-gkd-soft-vs-hard",
"status": "ready",
"manual_priority": 0,
"keywords": ["gkd", "distillation", "valid command", "task success"],
"text": "on-policy GKD made Qwen produce more valid shell commands: 7/19 → 13/19. strict task success stayed 1/4. this is exactly why agent evals need both soft syntax metrics and hard environment outcomes.\nhttps://huggingface.co/datasets/burtenshaw/agent-handover-august-2026"
},
{
"id": "gkd-families",
"sequence": 16,
"not_before": "2026-08-13",
"source_slug": "onpolicy-gkd-soft-vs-hard",
"status": "ready",
"manual_priority": 0,
"keywords": ["gkd", "qwen", "gemma", "teacher", "student"],
"text": "same GKD idea, different student/teacher families: Qwen improved command validity but not strict success; Gemma stayed at 0/4. teacher scale and falling loss did not rescue a weak task signal.\nhttps://huggingface.co/datasets/burtenshaw/agent-handover-august-2026"
},
{
"id": "qlora-behavior",
"sequence": 17,
"not_before": "2026-08-13",
"source_slug": "qlora-loss-vs-behavior",
"status": "ready",
"manual_priority": 0,
"keywords": ["qlora", "loss", "token accuracy", "behavior"],
"text": "12 QLoRA SFT steps improved eval loss 2.5109 → 2.1916 and token accuracy 0.4987 → 0.5351. then the behavior probe got worse: a requested one-word answer became a long refusal. loss moved, behavior regressed.\nhttps://huggingface.co/datasets/burtenshaw/agent-handover-august-2026"
},
{
"id": "pi-mono-transfer",
"sequence": 18,
"not_before": "2026-08-13",
"source_slug": "gemma-pi-mono-transfer",
"status": "ready",
"manual_priority": 0,
"keywords": ["pi mono", "gemma", "humaneval", "mbpp", "transfer"],
"text": "agent-trace SFT specialized unevenly. Gemma improved on HumanEval, about 0.7256 → 0.744, while MBPP fell 0.716 → 0.651. a model can look better on the benchmark nearest its traces and worse next door.\nhttps://huggingface.co/datasets/burtenshaw/agent-handover-august-2026"
},
{
"id": "awm-stress",
"sequence": 19,
"not_before": "2026-08-13",
"source_slug": "awm-session-stress",
"status": "ready",
"manual_priority": 0,
"keywords": ["agent world model", "openenv", "concurrency", "latency", "sessions"],
"text": "agent RL also fails below the model layer. a bounded OpenEnv stress run completed 24/24 sessions and 48/48 tool turns, but reset p90 was 13.226s while tool-call p90 was 0.006s. lifecycle dominates latency.\nhttps://huggingface.co/datasets/burtenshaw/agent-handover-august-2026"
},
{
"id": "recap",
"sequence": 20,
"fixed_date": "2026-08-31",
"not_before": "2026-08-31",
"source_slug": "campaign",
"status": "ready",
"manual_priority": 1000,
"keywords": ["recap", "lessons", "reproducibility"],
"text": "agent handover complete. dense environment supervision helped, small SFT worked, sparse RL rewards failed, and several soft metrics hid flat task success. all models, dashboards and reports stay here.\nhttps://huggingface.co/datasets/burtenshaw/agent-handover-august-2026"
}
]
}
|