qwen2.5-coder-1.5b-everywhere-exact-w8

<!-- Copyright (c) 2026 Anomly, Inc. All rights reserved. Author: Ry Bruscoe. -->

Anomly/qwen2.5-coder-1.5b-everywhere-exact-w8

One fine-tune. Every chip. Same bits. Every answer carries a receipt.

This is Qwen2.5-Coder-1.5B-Instruct fine-tuned on execution-verified coding data, with the training run, the served answers and the evaluation scores all carrying re-executable receipts, and served under an arithmetic profile that gives byte-identical outputs on x86 CPUs, NVIDIA CUDA and ARM.

What is in this repository

Numbers

Greedy decoding, pass@1, produced through the receipted endpoint (invar serve, exact profile):

One seed, greedy, 164 + 378 problems: differences of a point or two are noise. The fine-tune does not move the scores over the base; what it adds is the receipt chain, the identity board and the exact-8-bit training recipe. The base row, served in this exact 8-bit format, lands within a point of the bf16 scores in the Qwen2.5-Coder technical report.

Cross-hardware identity for the canonical prompt (Write a Python function that returns the n-th Fibonacci number, with a docstring.), 8 evaluations, every traced row of every layer plus logits:

The 0.5B sibling (Anomly/qwen2.5-coder-0.5b-everywhere-exact-w8, same recipe) gets the same board (the 1.5B ran on the 2 GB board from swap at 0.1 tok/s; the 0.5B is the one to actually use there):

For this 1.5B model the Blackhole leg is 3136/3136 rows, ACCEPT and the U200 leg is 3136/3136 rows, ACCEPT. Silicon legs re-execute the sampled rows with the exact accumulator in hardware; agreement is exact code equality, not tolerance. Full board with byte counts: identity/identity-board-2026-09-06.md.

How to verify

pip install anomly-invar[reexec]
invar verify evals/student15_fixed8/worldline.jsonl --model student15_fixed8-bposit8.gguf \
    --binary ./llama-cli --spot-check --units --reexec

--reexec replays each receipt from the weights and the token ids with an independent Python implementation (no llama.cpp code); invar-reexec (Go) does the same in a second language. The training chains are hash-chained per the Computation Receipts spec; the trainer that replays a segment (load the checkpoint at step N, re-run five steps, compare weights bit for bit) is not public yet, so today the chain is checkable and the replay is our claim with the replay receipts attached.

Training data

3,293 tasks written by a local teacher model, each kept only when its reference solution and an independently sampled second solution both pass the task's hidden tests in a sandbox; 8-gram decontamination against HumanEval and MBPP prompts; 12-gram self-dedup. The teacher's outputs carry a provenance record (SHA-256 of prompt and response, parameters), not a re-executable receipt: the teacher ran float vLLM. The student's chain is the re-executable part.

What this is not

Not smarter because of the arithmetic: any lift over the base comes from the data. Not confidential computing. Not a claim that receipts make answers correct. The Blackhole and FPGA verifier legs (if listed) re-execute sampled matmul rows, not the whole graph.

Licence

Apache-2.0 for the adapter, receipts and dumps. The base model is Apache-2.0 (Qwen). Contact: https://anomly.com/contact