Add 2 books to Qwen2.5-7B without it forgetting book 1 ($4,000)

We pay $4,000 for a training recipe that adds two new books to a Qwen2.5-7B-Instruct model that already knows one book, without losing it. Where we are. Book 1 is in the weights (our LoRA point, M1). Training book 2 on top works: 96 to 100 of 149 on our book-2 panel, up from 22 to 27. But the model loses book 1: only 12 of 161 key answers survive. Our best fix so far is a KL penalty to M1 on 300 book-1 lines. It keeps 103 of 161 and scores 87 to 93 on book 2. Scaling the same KL adapter down to 0.8 keeps 112 of 161 and scores 74 to 78 on book 2. Replay (older setup) stalled at 62% to 79% retention. Adapter merging, null-space projection and neuron targeting failed. We trained M1 and the book-2 points above on Claude-written notes, which you may not use. With clean data (GLM-5.2 teacher, no protection) book 2 reaches 53 to 69 and book 1 keeps 12 to 18 of 161. Other numbers share one fixed answer setup. Book-2 scores use a calibrated LLM judge. The task. Train book 2, then book 3, on top of M1, with one recipe and no manual steps. You hand us a frozen package. We run it ourselves on our GPU with our 3 seeds and score it on panels you never see. What we pay for. All lines at once, averaged over our 3 seeds: - Book 2 is learned: at least 72 of 149 on the book-2 panel, and at least 20 above M1 with the 95% interval above zero. - Book 1 is kept: at least 130 of 161 key answers (18 over our 0.8 point; our best full point is 103), no seed below 112, and at least 75% of the book-1 panel questions M1 gets right (79 of 150, same judge) stay right, no seed below 70%. - Honesty does not get worse: on 60 false-premise questions, at most 4 more made-up answers than M1 (M1 makes 32). At most 2 of 30 refusals on true questions of the same shape and at most 2 of 15 on everyday questions, on every seed. - General skills stay, in our fixed lm_eval setup (commands shared at contract start): IFEval prompt-level strict at least 70.35, GSM8K 5-shot strict-match at least 74.21, MMLU 5-shot on our 2,000-question subset at least 70.00. Drops below the reference (base for IFEval and GSM8K, 71.35 and 75.21; M1 for MMLU, 71.00, since M1 itself is 2.95 below base there) add up to at most 1.5, and a gain on one test does not cancel a drop on another. If we cannot rebuild our Oct 14 stack, we re-measure base and M1 next to your model and apply the same rule. - Cost, in every run: at most 8 GPU hours per book on one 48 GB card, all steps included, and at most $5 of teacher API per book. - After book 3: book 3 learned (at least 75% of what our own run with no protection scores on the book-3 panel; the number is fixed in writing before milestone 3 starts), at least 19 of 140 above your book-2 point with the 95% interval above zero, book 1 still at 130 of 161 or more with no seed below 112, at least 75% of the book-2 answers that were correct stay correct with no seed below 65%, and the honesty, skills and cost lines still hold. Exact definitions are on our terms page: https://x10-terms.13-140-178-130.sslip.io/. Its hash goes into the contract and does not change. Rules. - Base is Qwen2.5-7B-Instruct, revision a09a3545. One checkpoint and one way to run it. - You may add modules (experts, memory layers, heads) and freeze old ones. The model has to pick what to use by itself, on a plain question without the book name. Added modules come with full loader source and must load in our lm_eval setup, as hf or through your adapter. - No RAG, no search at answer time, no adapter switching by book. - No text from Claude, GPT, Gemini or any other closed model in the training data. Allowed teachers: Qwen2.5-7B-Instruct itself, open-weight models whose license allows training on their outputs, and GLM-5.2 through our key. An open-weight teacher needs our yes before you use it. We answer within two days. - Code or data with no license or a non-commercial (NC) license fails the clean-data line, so Memory Layers code (CC-BY-NC) is out. - Our test questions, including the IFEval, GSM8K and MMLU sets we score, never go into training. We scan your training data for overlap. Money. Three fixed-price milestones on Upwork. We fund each one in escrow before its work starts, and we pay only for results we measure on our side. 1. $500, first result: frozen package, a plan with your predicted numbers, and one book-2 run on our GPU (one seed) that already clears every line above for the book-2 step, except general skills. 2. $1,500, book 2: all book-2 lines on our 3 seeds. 3. $2,000, book 3: all lines after book 3 on our 3 seeds, plus the full package and a written report. It opens only after a judge passes our check on the book-3 panel (ours passed on books 1 and 2, failed on book 3). If no judge passes within 30 days after the milestone 2 verdict, it is cancelled and neither side owes anything. Our re-run of your package with the same seed must give the same training data wherever generation is deterministic. If something breaks on our side, we re-run up to twice at our cost. Before each verdict we re-check our judge on 80 fixed gold pairs and recalibrate it at our cost if it drifts. If a score lands inside the judge's uncertainty band, we re-judge up to twice at our cost. If it is still inside, the line is not reached. After the milestone closes, you get the answers and labels for that line and our scoring code. The milestone cannot be retried on that panel. You can ask for an independent recount. We pay up to $150 for it. Upwork takes its own fee from your side. What you get when the contract starts: the English source text of books 2 and 3 split into units, M1 weights, our answer-generation script, a description of every panel, a short brief of what we tried and scored, GPU time on our pod within a fixed budget, and our judge on up to 600 of your own dev questions per milestone. Answer the three questions below, numbered. We decline proposals without a public artifact